Lecture Notes in Computer Science 1603

Edited by G. Goos, J. Hartmanis and J. van Leeuwen

Springer
Berlin
Heidelberg
New York
Barcelona
Hong Kong
London
Milan
Paris
Singapore
Tokyo
Jan Vitek Christian D. Jensen (Eds.)

Secure
Internet Programming
Security Issues
for Mobile and Distributed Objects

^M Springer
Series Editors
Gerhard Goos, Karlsruhe University, Germany
Juris Hartmanis, Cornell University, NY, USA
Jan van Leeuwen, Utrecht University, The Netherlands

Volume Editors
Jan Vitek
University of Geneva, Object Systems Group
CH-1211 Geneva 4, Switzerland
E-mail: [email protected]

Christian D. Jensen
Trinity College Dublin, Department of Computer Science
O'Reilly Institute, Dublin 2, Ireland
E-mail: [email protected]

Cataloging-in-Publication data applied for

Die Deutsche Btt>Uothek - CIP-EiiiheitsiHifiiahnie

Seoire Interact prsgnnuBtaf: security issueaformobile and distributed
otigects / Jan Vitek; Christian D. Jensen <cd.). - Berlin ; Heiddberg; New
Yoik; Baioel«ia; Hong Kcmg; London; Milan; Paris; ^Bgapcm; Tokyo
: Sprii^r, 1999
(Lecture notes in c o n f e r scioice ; Vol. 1603)
ISBN 3-540-66130>1

CR Subject Classification (1998): C.2, D.1.3, D.1.5, D.2.4, H.4.3,1.2, K.4.4,

K.5.1,K.6.5
ISSN 0302-9743
ISBN 3-540-66130-1 Springer-Verlag Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,1965,
in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
liable for prosecution under the German Copyright Law.

© Springer-Verlag Berlin Heidelberg 1999

Printed in Germany
Typesetting; Camera-ready by author
SPIN: 10704834 06/3142 - 5 4 3 2 1 0 Printed on acid-free paper
 Preface

This book is a collection of articles on security in large-scale open distributed systems

and, in particular, issues pertaining to distributed and mobile object technologies. The
role of open systems is to provide an infrastructure for assembling global applications
out of software and hardware components coming from multiple sources. Open systems
rely on publicly available standards to permit heterogeneous components to interact.
The Internet is the archetype of an open distributed system. Standards such as HTTP,
HTML, and XML together with the widespread adoption of the Java language are the
cornerstones of many distributed systems. Distributed object technologies such as RML
CORBA, and DCOM, as well as more ambitious mobile technologies (mobile agents,
ambients, etc.) offer a large palette of approaches to building those systems.
While the benefits of these technological solutions are undeniable, they raise le-
gitimate security questions. What kind of trust can we have in a system composed of
components belonging to different organisations and used by users from all comers of
the world? Open systems lack a common trusted computing base, and often have mini-
mal authentication and authorisation frameworks. The possibility of moving code from
site to site raises the stakes even higher. We must be able to establish trust among users
and programs, and to specify policies that will regulate the use of resources and the
spread of information. We must find ways to integrate untrusted downloaded programs
with audited local code so that security remains meaningful. We must investigate means
to systematically enforce security properties in programming languages for those sys-
tems. And, if everything else fails, we must be able to diagnose when policies have been
breached and pinpoint the guilty party.
These are a few of the challenges facing us today. The papers in this collection are
a snapshot of this very active research area.
The idea for this book arose from a number of discussions and presentations given
at two workshops: the ECOOP Workshop on Distributed Object Security (EWDOS) and
the Mobile Object Systems: Secure Internet Mobile Computations (MOS98) workshop
held in June 1998 in conjunction with the European Conference on Object-Oriented
Programming in Brussels. Both workshops were reviewed and the authors of the best
papers of each invited to contribute to this volume. In addition to these core papers, a
small number of classic papers in the field were reprinted. The outcome is a well bal-
anced collection of papers that explore many of the essential questions outlined above.

Overview

The book is organised in three parts: (I) Foundations, (II) Concepts, and (III) Imple-
mentation, followed by an appendix.
Part I of the book contains chapters giving background and dealing with funda-
mental issues in trust, programming, and mobile computations in large scale open dis-
tributed systems. The paper by Swarup and F^brega is a comprehensive study of trust in
open distributed systems. The paper by Abadi discusses abstractions for protection and
 VI

the correctness of their implementation. The paper by Ancona et al. analyses the use
of reflection to integrate authorisation mechanisms into an object-oriented system. The
paper by Cardelli discusses the difficulties of computing with mobility and proposes a
unified framework to overcome these difficulties based on mobile computational am-
bients. The paper by Hennesy and Riely studies the type safety properties of mobile
code in an open distributed system. The paper by De Nicola et al. describes the secu-
rity mechanisms of the programming language KLAIM, that is used to program mobile
agents. The paper by Leroy and Rouaix formulates and proves security properties that
well-typed applets possess and identifies sufficient conditions for the execution envi-
ronment to be safe.
Part II contains descriptions of general concepts in security in open distributed sys-
tems. The paper by Blaze et al. describes the use of trust management engines to avoid
the need to resolve "identities" in an authorisation decision. The paper by Aura dis-
cusses the advantages and limitations of delegation certificates in access control mech-
anisms. The paper by Brose proposes a new fine grained access control model for
CORBA based on views. The paper by Tschudin introduces the concept of apopto-
sis (programmed death) in mobile code based services. The paper by Yee discusses the
problem of ensuring confidentiality and integrity for mobile agents. The paper by Roth
describes how two co-operating agents, executing on different machines, can be used
to protect the confidentiality and integrity of the individual agent against tampering.
Part III contains papers detailing implementations of security concepts in open dis-
tributed systems. Most of the papers in this part also introduce new security concepts,
but devote a large portion to a particular implementation of these concepts. The paper
by Jaeger describes the use of role based access control policies in configurable sys-
tems and its implementation in the Lava Security Architecture. The paper by Grimm
and Bershad discusses secure execution of possibly untrusted extensions in the SPIN
extensible operating system. The paper by Jones describes a simple and efficient way
of interposing code between user programs and the underlying system's interface. The
paper by von Eicken et al. discusses the inadequacy of object references for access
control and describes an implementation of capabilities in the J-Kernel. The paper by
van Doom et al. describes the implementation of secure network objects, which is an
extension of Modula-3 network objects. The paper by Edjlali et al. describes an ac-
cess control mechanism in which access is granted based on the history of interactions
with the requesting principal. The paper by Alexander et al. discusses security issues
in active networks, and the solutions that have been implemented in the Secure Ac-
tive Network Environment. The paper by Hulaas et al. discusses the problem of mobile
agent interactions in an open environment. The paper by Wilhelm et al. describes how
trusted tamper resistant devices can be used to ensure the integrity of mobile agents.

Acknowledgements

We would like to thank the members of the programme committees of the two work-
shops. For EWDOS, they were: George Coulouris, Leendert van Doom, Li Gong,
Daniel Hagimont, Trent Jaeger. For MOS98, the committee included: Martin Abadi,
Brian Bershad, Ciaran Bryce, Luca Cardelli, Giuseppe Castagna, Robert Gray, Leila
 VII

Ismail, Dag Johansen, Eric Jul, Doug Lea, Christian Tschudin, Dennis Volpano. Fur-
thermore, we wish to thank: Vinny Cahill, Daniel LeMetayer, Tommy Thorn, Hitesh
Tewari, and Mary Ellen Zurko for additional reviewing, and Alfred Hofmann from
Springer-Verlag for his support in getting the volume published.

February 1999 J. Vitek and C. D. Jensen

 Table of Contents

I Foundations

Trust: Benefits, Models, and Mechanisms

(Vipin Swarup and Javier Thayer Fdbrega) 3
Protection in Programming-Language Translations
(Martin Abadi) 19
Reflective Authorization Systems: Possibilities, Benefits, and Draw-
backs (Massimo Ancona, Walter Cazzola, and
Eduardo B. Fernandez) 35
Abstractions for Mobile Computations (Luca Cardelli) 51
Type-Safe Execution of Mobile Agents in Anonymous Networks
(Matthew Hennessey and James Riely) 95
Types as Specifications of Access Policies
(Rocco De Nicola, GianLuigi Ferrari, and Rosario Pugliese) 117
Security Properties of Typed Applets (Xavier Leroy and
Frangois Rouaix) 147

II Concepts

The Role of Trust Management in Distributed Systems Security

(Matt Blaze, Joan Feigenbaum, John loannidis, and
Angelos D. Keromytis) 185
Distributed Access-Rights Management with Delegation Certifi-
cates (Tuomas Aura) 211
A View-Based Access Control Model for CORBA (Gerald Brose) 237
Apoptosis — the Programmed Death of Distributed Services
(Christian Tschudin) 253
A Sanctuary for Mobile Agents (Bennet S. Yee) 261
Mutual Protection of Co-operating Agents (Volker Roth) 275
Ill Implementations

Access Control in Configurable Systems (Trent Jaeger) 289

Providing Policy-Neutral and IVansparent Access Control in
Extensible Systems (Robert Grimm and Brian N. Bershad) 317
Interposition Agents: Transparently Interposing User Code at the
System Interface (Michael B. Jones) 339
J-Kernel: A Capability-Based Operating System for Java
(Thorsten von Eicken, Chi-Chao Chang, Grzegorz Czajkowski,
Chris Hawblitzel, Deyu Hu, and Dan Spoonhower) 369
Secure Network Objects (Leendert van Doom, Martin Abadi,
Mike Burrows, and Edward Wobber) 395
History-Based Access Control for Mobile Code (Guy Edjlali,
Anurag Acharya, and Vipin Chaudhary) 413
Security in Active Networks
(D. Scott Alexander, William A. Arbaugh, Angelas D. Keromytis,
and Jonathan M. Smith) 433
Using Interfaces to Specify Access Rights (/. Hulaas, A. Villazon,
and J. Harms) 453
Introducing Trusted Third Parties to the Mobile Agent Paradigm
(Uwe G. Wilhelm, Sebastian Staamann, andLevente Buttyan) ... 469

IV Appendix

List of Authors 493

 Parti

Foundations
 Trust: Benefits, Models, and Mechanisms

Vipin Swarup and Javier Thayer Fabrega

Abstract. Trust provides many practical benefits in open distributed

systems. It enables cooperation between pairs of entities, provides a safe
and inexpensive ba;Sis for lowering access barriers to secured resources,
and facilitates complex transactions among multiple entities. In this pa-
per, we describe a wide range of notions and aspects of trust in open
systems, with particular focus on trust between autonomous entities.
We argue for computational models and mechanisms that enable trust
between entities to be produced, manipulated and degraded and we sum-
marize the state-of-the-art from this perspective.

1 Introduction
Trust is a fundamental concept in computer security. However, while many cen-
tral concepts in computer security such as privacy and integrity now have com-
monly accepted definitions, trust remains an ambiguous term. There are nu-
merous notions of trust in the computer security literature, including trust as
assurance in the correct and secure functioning of software, computer systems,
and legal systems; trust as belief in the benevolent, honest, competent, and pre-
dictable behavior of autonomous agents (human or software); and trust as a
tendency to depend on others. Different kinds of trust often satisfy different
properties; for instance, some kinds of trust are transitive (e.g., trust in the con-
text of network administration tools such as SATAN [II]) while others are not
(e.g., trust in the context of name-key binding mechanisms such as PGP).
Trust-based security architectures are widely deployed. For instance, in net-
work administration, trust is a relationship that exists between an entity man-
aging local resources and a remote client, whenever that client can access lo-
cal resources without local authentication or authorization (e.g., .rhost files in
Unix) [11]. In network security, a prevalent paradigm is to partition the network
into regions and to use firewalls to filter network packets between regions; the fil-
tering policies implemented by the firewalls are based on the trust relationships
between regions [16]. In distributed system security, participants use identity
certificates to identify each other; they need to trust other participants to is-
sue good certificates and to recommend reliable third parties for accomplishing
specific tasks, including signing certificates [14, 9]. In mobile agent security, a
mobile agent may be given access to resources at a host based on the host's
level of trust in the agent; this trust can depend on many factors including the
endorser of the agent, the hosts that the agent has visited in the past [12, 2],
and the access requests that the agent has made in the past [8].
Trust is important due to the many practical benefits it provides. For in-
stance, trust serves many important roles in agent societies: it is central to
agents engaging in cooperative activities; it provides an inexpensive (though of-
ten imprecise) basis for lowering expensive access barriers between agents; and
it enables agents to form clusters within which complex transactions have a high
likehhood of success. We need computational models of trust that describe how
trust between autonomous agents can be produced, manipulated, and degraded.
Agents can develop trust in each other by using mechanisms that implement
those models [21, 27]. This will enable software agents to trade in unfamiliar
environments with unfamiliar trading partners by relying on trust that is built
remotely across computer networks and appears essential for agents to play an
important role in distributed electronic commerce.
In this paper, we elaborate on the various distinctions on trust that are
mentioned above, drawing from a vast literature on trust in the social sciences
and in computer security. In Section 2, we present a classification of the various
meanings of trust. Although we consider the full range of meanings of trust, our
focus is on trust between agents and not on system trust or dispositional trust.
In Section 3, we motivate the importance of trust by describing its central role in
three significant areas: cooperation, lowering of access barriers, and clustering. In
Section 4, we describe how trust can be produced, manipulated, and degraded.
In Section 5, we summarize graph models of trust propagation that have been
proposed in the hterature. Section 6 concludes this paper.

2 Classification

Trust is not a well-defined concept - numerous definitions of trust can be found in

the literature. McKnight and Chervany [23] have proposed a classification system
that distinguishes between several kinds of trust. Although this classification is
not comprehensive, it does cover some of the more important forms of trust
and is useful in describing different aspects of a computational trust model. We
summarize their classification scheme below; we have grouped their trust classes
into three broad categories: system trust, entity trust, and dispositional trust.

System Trust: "The extent to which an entity believes that proper system
structures are in place to enable it to anticipate a successful future en-
deavor." [23] Safeguards such as regulations, guarantees, and security mecha-
nisms reduce the potential negative consequences of trusting behavior. Stabi-
lizing intermediaries (e.g., insurance companies) reduce uncertainty and risk
(e.g., financial risk). Most work in information systems security is aimed
at managing risk, thus increasing system trust and enabling entities to feel
more secure in trusting others.
Entity Trust: Entity trust includes all trust relationships between two or more
entities. These relationships can be further categorized as follows:
Trusting Belief: "The extent to which an entity beheves that another en-
tity is willing and able to act in the former's best interests." [23] The
belief can be based on a variety of attributes including the other's benevo-
lence, honesty, competence, and predictabihty. Assessing these attributes
 directly is an inherently subjective process and is typically handled by
procedural mechanisms.
As an example, consider public key certification schemes (e.g., PGP and
X.509) that use identify certificates to bind keys to names. The schemes
include computational trust models [14,10] that enable the holder of a set
of certificates to decide whether to believe a particular name-key binding.
This decision may be based on the holder's belief in the trustworthiness
of the issuers of the certificates. We will examine several such models in
Section 5.
Trusting Intention: "The extent to which an entity is willing to depend
on another entity in a given situation with a feeling of relative security,
even though negative consequences are possible." [23] For instance, if a
trust metric for the PGP trust model yields some trust value for the
issuer of an identity certificate, the holder of the certificate must decide
whether the trust value is sufficient for it to believe the name-key binding
stated in the certificate. This decision function represents the trusting
intention of the holder.
Trusting Behavior: "The extent to which an entity depends on another
entity in a given situation with a feeling of relative security, even though
negative consequences are possible." [23]. For instance, a .rhost file on a
Unix host represents unconditional trusting behavior. If a user joe on
host orion creates a .rhost file that contains the name of host p l u t o ,
then any user authenticated as joe on p l u t o can access joe's account
on orion without further authentication. Thus joe must trust p l u t o
to enforce appropriate security safeguards to protect joe's account on
orion.
Trusting belief, intention, and behavior are described as relations between
two entities. However, these same notions can describe trust relationships
between multiple entities. McKnight and Chervany [23] identify the following
kind of multi-entity trust as important and prevalent:
Situational Decision to Trust: "The extent to which an entity intends
to depend on other entities in a given situation." [23] For instance, a
user may intend to depend on the system administrators (to recover her
files) in the event that her files get corrupted.
Dispositional Trust: "The extent to which an entity has a consistent ten-
dency to trust across a broad spectrum of situations and entities." [23] This
kind of trust may be viewed as a policy or strategy which produces trusting
intentions. It could arise because the entity believes others to be trustworthy
in general, or because it expects to obtain better outcomes by trusting others
irrespective of whether they are trustworthy. In computational trust models,
dispositional trust is manifested by threshold values. For instance, an entity
may enter a dependency relationship with another entity if its trusting belief
expectation in that entity exceeds its dispositional trust threshold.

In the remainder of this paper, we explore different aspects of entity trust

between two or more autonomous agents. Although we do not consider the sub-
categories of entity trust in this paper, a real-world computational model of
entity trust must account for these distinctions.

3 Benefits

System trust plays an important role in enabling and sustaining cooperation be-
tween agents. Entity trust between two agents reduces the costs of identification
and access control between the agents, while entity trust among multiple agents
reduces the costs of complex multi-agent transactions. These benefits underscore
the importance of trust in distributed agent systems.

3.1 Enabling Cooperation

Agents will be willing to cooperate (e.g., engage in e-commerce transactions)

only if they trust that the system satisfies certain properties (e.g., their credit
card numbers are protected). Williams [32] defines cooperation as follows: "Two
agents cooperate when they engage in a joint venture for the outcome of which
the actions of each are necessary, and where a necessary action by at least one
of them is not under the immediate control of the other". Under this definition,
if two agents cooperate, then at least one of them depends on the other. For
instance, in an e-commerce transaction, the buyer of a product depends on the
seller to dispatch the product, while the seller depends on the buyer to dispatch
payment; in this situation, the buyer and seller depend on eaxih other.
Williams [32] asserts that cooperation can arise only in the presence of two
motivations (in the same society, not necessarily in the same agent): agents must
be motivated to enter into dependent positions, and agents must be motivated
not to defect if they are in a non-dependent position. Dependent agents need
some degree of assurance that non-dependent agents will not defect, and coop-
eration will happen only if this assurance is to some extent well based. System
trust plays an important role in providing this assurance, thus enabling and
sustaining cooperation among agents.

3.2 Lowering Access Barriers

A major goal of security is the protection of resources from undesired access. In

practice, this is achieved by creating barriers around resources. A barrier may
constrain every access to a resource, or it may constrain access into a domain
from which a number of resources may be accessed without further constraints.
In the physical world, permission to traverse a barrier may depend on the identity
of the requestor and perhaps other considerations such as payment of a fee. The
identity of the requestor is determined by a picture ID such as a license or a
passport and the adequacy of the fee is determined by characteristics of the
paper used for payment such as treasury watermarks and the denomination of
the bill.
 Traversing access barriers can be tedious, consume time and resources, and
discourage commerce and productive behavior. For example, Fukuyama [13]
mentions the possibihty of patrons sneaking out of restaurants without paying
(and thus having illegal access to food). To prevent this, owners could require
cash deposits before seating patrons. Conversely, patrons who fear being cheated
could demand certificates of competence before entering restaurants. The result-
ing war of mistrust would likely lead to restaurants going out of business and
the public losing the pleasures of dining out.
Similarly, U.S. stores consider a current driver's license from any U.S. state
as valid identification for cashing checks, thus trusting the states' ability to issue
valid name-picture certificates. The cost to the stores and the inconvenience to
shoppers would be too great if each store had to issue its own identity certificates.
As another example, consider a person who wishes to purchase a round-the-
world air ticket at a low price. She locates a travel agent who can provide her with
a ticket at a substantially lower price than the airline's lowest published fare. The
travel agent obtains the ticket from a consolidator who purchases blocks of unsold
seats from the airline at discount prices. This is a complex transaction involving
four agents and multiple trust relationships. The prospective traveler must trust
the travel agent to provide her with a valid ticket, the consolidator to pay the
airhne before the flight, and the airline to provide her with transportation. The
travel agent must trust the consohdator to obtain seats. These trust relationships
reduce the need for safeguards (such as legal contracts) that would increase the
cost of the transaction.
These examples provide analogies to why trust is needed in computer security.
In particular, trust reduces the cost of access control. It is a basis for granting
privilege by lowering access barriers selectively [5]. Further, trust reduces the cost
of identification. It provides an identification infrastructure which decentralizes
the creation of identity certificates [4].

3.3 Creating Communities

Dual to the notion of trust as a means of reducing the cost of access barriers be-
tween individuals is that of trust as an indication of the willingness of individuals
to associate. Fukuyama [13] defines sociability as a measure of the propensity to
form associations (independent of the efficacy of those associations). Human as-
sociations can be classified by the nature of their origins, for instance, those that
originate from family relationships or those that are created by the initiative of
some authority such as the state. Spontaneous sociability, that is propensity to
form associations derived from sources other than kinship or the state, is a pre-
requisite for the formation of a large class of organizations. One of Fukuyama's
main theses is that spontaneous sociability is a cultural characteristic directly
related to the levels of trust in a society.
These concepts are useful in understanding how culture affects economic
institutions, particularly cooperative behavior in organizations and the structure
of business ownership. For instance, a large organization may require processes
that involve many employees and have no central point of control. For these
processes to succeed, employees must be motivated to initiate and participate
in activities which might fail if other employees in the workflow do not perform
competently. This motivation requires trust in the capabilities and motivations
of other employees in the organization. This trust goes beyond system trust or
pairwise trust.
However, these concepts also provide a framework for understanding associ-
ations in any community, including software agents. In order for this approach
to be useful in explaining the growth and dynamics of associations, there has to
be some account of counterbalancing forces, such as diminishing economies of
scale, which limit the size of associations.
At a more abstract level, the propensity to form associations within some set
of individuals is a measure of the capacity for clustering. There are many exam-
ples of clustering in physics, for instance, Ising models to explain spontaneous
ferromagnetism. To produce an economic model along similar lines, note that a
successful business must be very effective in performing tasks that require many
transactions. This suggests a transaction model for certain kinds of social activi-
ties. In this model, we associate a transaction T{ki, ^2, a) with individuals ki, k2
and activities a. The transaction T{ki,k2,a) may succeed or fail. A cluster is a
set of individuals in which suites of transactions between individuals in the set
are likely to succeed.

3.4 Example: Mobile Computations

Mobile computations offer a pai-adigm for computing in the presence of network

access barriers. For instance, consider two agents that are separated by a firewall
and that wish to communicate using a custom protocol that is blocked by the
firewall. One common solution to this problem is for the agents to tunnel through
the firewall—the agents encapsulate their communications as data within a pro-
tocol that the firewall accepts, thus bypassing the filtering constraints of the
firewall. Mobile computations offer another solution whereby an agent migrates
across the firewall. The firewall may enforce a filtering policy on agent migration;
thereafter, the agents may communicate without restriction.
However, when mobile computations cross the security domain boundaries
guarded by firewalls, they may cause trust relationships to change in complex
ways [2, 12]. For instance, a host's trust in a software agent may vary depending
on whether the agent has migrated through other potentially hostile hosts. As
another example, a system may no longer assume that all hosts within a security
domain are trusted equally (since hosts may be mobile), or that all code installed
in a domain is trusted (since code may be mobile).
Further, mobile agent systems involve complex trust relationships among
hosts, program code, agents, and users. For instance, agent routing policies may
prevent agents from migrating to untrusted servers. However, since agents are
routed and executed by servers, the agents must trust some servers to enforce
their routing policies. The servers, in turn, may have to trust incoming agents
to respect their security policies.
 The problems are exacerbated in wide-area networks where entities have no
knowledge of most other entities or of their trust relationships. Traditional static
assumptions about trust are inappropriate—these systems need trust models
that enable complex trust decision rules to be established dynamically. As de-
scribed earlier, this dynamically determined trust can play an important role
in enabling cooperation between agents, lowering access barriers between agents
and hosts, and supporting the clustering of agents into communities.

4 Operations
Entity trust between autonomous agents can be produced by operations such
as multiple successful interactions, manipulated by operations such as propaga-
tion and substitution, and degraded by operations such as natural decay and
malicious attacks.

4.1 Production

Trust can be established directly between two or more agents in a variety of

ways [24, 18].

Swift trust is based on a series of hedges in which agents behave in a trusting

manner but also hedge to reduce the risks of betrayal. Blind trust is an
extreme case of this where an agent trusts without hedges because that is
the best choice in its situation.
Deterrence-based trust is based on a credible threat of punishment for cheat-
ing or failing to cooperate. For instance, when there are repeated interac-
tions, cheaters are deterred by the possibility of being punished in the fu-
ture. Audit and audit reduction mechanisms are examples of this, where
the mechanisms raise a credible threat of being detected and punished for
circumventing the security policy of a system.
Knowledge-based trust is based on sufficient knowledge of other agents' dis-
positions so that their behavior can be predicted reliably. For instance, in
history-based access control models [8], an agent's access rights depend on
the access requests made by the agent in the past. In some probabilistic
models [22], an agent's trust in another agent depends on the number of
successful interactions the former has had with the latter.
Identification-based trust is based on common preferences where agents have
taken on the needs and desires of others as personal goals and act in ways
that consider joint gains. For instance, a firewall establishes a trust domain
within which all agents trust each other based on some common identifica-
tion such as being agents of employees of the same company.
Social trust is based on social relationships between people, for instance, neu-
trality and social status. Social interactions help define the identities and
social worth of people, and these in turn influence social relationships. So-
cial trust is based on emotions in contrast with the previous operations
 10

which are based on rational thought. In agent systems, social trust appears
in models where one's trust in an agent is based on one's social trust in the
principals that have signed (and hence endorsed) the agent.

Current computational trust models assume that direct trust valuations are
given as input and they describe how to compute indirect trust (see Section 4.2).
Establishing and assessing direct trust is considered to be subjective and is
outside the scope of these models. By contrast, agent-based systems cannot rely
on manual trust valuations and need technical mechanisms for producing trust
directly.

4.2 Manipulation

An agent may not be able to establish trust in another agent directly. The agent
can then use mechanisms that obviate the need to establish the trust [26], or
mechanisms to establish trust indirectly (that is, involving other "third-party"
agents).

Trust substitution mechanisms alter the trust relationships that would oth-
erwise be required between agents. These include technical solutions that
eliminate one's need to trust an agent, substituting it with trust in a third
party instead. For instance, in the physical world, the trustworthiness of a
tamper-resistant smartcard depends on the manufacturer of the card and
is independent of the entity that physically hosts the card. If the card is
used for electronic commerce, then participants in transactions need only
trust the manufacturer of the card and not necessarily its owner. Mobile
cryptography [30] provides similar fimctionality for software agents.
Trust propagation mechanisms propagate trust from one principal to another.
This kind of trust is known a^ indirect trust. For instance, professional or-
ganizations and recommendation services impart trust by virtue of their
reputations, and insurance companies impart trust by increasing the (finan-
cial) predictability of outcomes. In all these cases, trust flows from reputable
institutions to other principals.
Propagation mechanisms typically combine different trust valuations from
different principals into new trust valuations. Certain kinds of trust are
transitive while others are not. Trusting behavior (see Section 2) is often
transitive. For instance, if a UNIX host venus has a trusting relationship
with host p l u t o via a .rhost file, and p l u t o has a similar relationship with
host orion, then venus also has a trusting behavior relationship with orion.
Trusting beliefs are often not transitive. For instance, if Nisha trusts Ignacio
to make good recommendations, and Ignacio trusts Williaan to make good
recommendations, then Nisha does not necessarily trust William. Most work
on formal computational trust models deals with non-transitive propagation
trust; the aim of these models is to address the name-key binding problem
of identification and authentication. Section 5 presents a summary of such
models.
 11

4.3 Degradation

Trust is time dependent. In fact, without reinfoicing interactions, trust may

decay naturally over time. Tiust may also degrade during propagation. Further,
trust is subject to attack in that princijials may subvert some of the elements that
support a network of trust. For instance, a human principal can be physically
attacked and its keying material stolen.
Trust can also be destroyed by systematic operations and techniques, usually
in situations where compromise is suspected. This may protect honest agents
from making decisions on biased trust data. Key revocation schemes are an
example of this—a key revocation certificate is issued when a key may be com-
promised; the certificate destroys other agents' trust in the revoked key.

5 Models

The social science literature contains numerous models of trust production while
the computer security literature contains numerous models of trust propagation.
Trust production is a subjective process for which adequate computational mod-
els have yet to be developed. Thus, in this section, we examine formal models of
trust propagation only and we describe some of their limitations. The axiomatic
(or mathematical) models of trust discussed here are graph models in the sense
that the mathematical structure used to model trust is a finite directed graph;
we do not consider models based on belief logics [25, 7, 1, 19].
In graph models, the nodes of a graph represent principals and the directed
edges of a graph represent trust relationships between principals. There may
be additional structure associated with a grai)h, such as labels on the nodes
and edges. The graphs are used to pioj^agate beliefs about statements (e.g.,
"the key P belongs to Nisha"). The trust graphs are extended with nodes that
represent statements and with edges from principal nodes to statement nodes
that represent certificates (i.e., signed statements). These axiomatic models are
useful to the extent they ahow deduction and calculation using unambiguous
rules. A given axiomatic model still needs justification to determine whether it
says anything useful about the real world.

5.1 Simple Graph Models

The simplest kind of graph model is given as follows:

1. The graph consists of two disjoint sets of nodes: A set of principals (keys)
and a set of statements (e.g., name-key bindings).
2. Directed edges are of two kinds:
— An arrow fc —>fc'means k signs a certificate saying "k trusts k' to issue
certificates with valid statements only". The trust represented by this
edge has been referred to in the hterature as recommendation trust ([3]).
There is a similar concept in [20] referred to as delegation.
 12

Nisha

Fig. 1. Simple trust graph

— An arrow k -^ s means k signs a certificate saying s. Again there is a

similar notion in [20].

Figure 1 contains a simple trust graph in which there are seven principals
A,B,C,D,E,F,H and three statements P: Nisha, Q: Ignacio, R: William. The
statement P: Nisha asserts that the key P belongs to Nisha. The edge C ~»
P: Nisha represents a certificate issued by C that says P: Nisha. The edge B ^ C
represents a certificate issued by B that says "B trusts C to issue certificates
with valid statements only".

5.2 Trust Decisions

A graph model provides a graphical picture of the keys, statements, certificates,
and trust relationships in a community at a particular moment. It is important
to note that the trust graph by itself provides no information about which are
the valid statements nor does it indicate whether a principal should believe a
statement. The model must contend with the fact that trust can be unfounded
since, in the real world, there are principals who issue false certificates for fun
and profit.

Definition 1. A state of a trust graph is a svhset G,, of the set of statements,

intended to denote the set of valid statements. A principal k is fraudulent with
respect to Gy if there is an edge k -^ s 0 G„.
Figure 2 illustrates a state of a trust graph in which there are two valid
statements P: Nisha, Q: Ignacio and a fraudulent principal F which certifies the
invalid statement R:Willicun.

Deflnition 2. Let Q he the set of trust graphs. A trust decision rule (3 associated
to Q is a boolean function {G,k,s) i-> 0{G,k,s) defined whenever G is a graph
in Q, k is a key in G, and s is a statement in G.

Figure 3 illustrates a trust decision rule in which the rule is /?(G, k, s) = t r u e

only when there are two independent corroborating trust paths from k to the
 13

>:Nisha

Q:Ignacio

R:William

Fig. 2. State of a trust graph

Accept P: Nisha
B C

D.
7^
P:Nisha
H
E Q:Ignacio

RiWilliam

Fig, 3. Example of a trust decision rule

statement s. Thus, according to this rule, the keys A,B,D are the only ones to
accept the statement P: Nisha. Moreover this rule rejects ail other statements.
Notice that under this trust decision rule, nodes C, F, H in the example should
not accept any statements, including the statements they certify; these nodes
may accept statements based on other models of trust such as a model of direct
trust. The hterature describes several trust decision rules including rules where
/3(G, k, s) = t r u e only when there is a path from k to s whose length is smaller
than some threshold [33, 14], or when the number of bounded node-disjoint paths
from k to s exceeds some threshold [28].
Notice that a trust decision rule depends only on the graph and is indepen-
dent of the trust state. Moreover, as defined above, the value of /3(G, k, s) may
depend on the structure of the entire graph, including parts of the graph preced-
ing the node k along an oriented path. However, a key k should decide trust not
on who trusts k but on whom k trusts. To make this notion precise we define a
trust "interval": Given a statement node s and a key node k of G, G[k, s] is the
set of nodes that lie on some oriented path starting at k and ending at s. Notice
that k, s are both nodes in G[k,s]. Moreover, s is the only statement node in
G[k,s\.
 14

Definition 3. An interval trust graph is one of the form G\k,s\.

We make the following two assumptions on trust decision rules;

1. If G,G' are isomorphic via an isomorphism which maps k M- k' and s H->
s', then P{G,k,s) = fi{G',k',s'). That is, 0{G,k,s) depends only on the
structure of G and not on the labels on its nodes.
2. P(G,k,s) — P{G[k,s],k,s). That is, p{G,k,s) depends only on the interval
structure of G between k and s
The following easily proved fact describes all the trust decision rules:
Any boolean function (ji on interval graphs defines a trust decision rule
by l3^{G,k,s) = 0(G[fc,s]). Moreover, any trust decision rule is of this
form.

5.3 Comparison of Trust Decision Rules

One way of evaluating a trust decision rule is by computing the proportion of
keys that erroneously accept or reject statements. We formalize this concept
using a notion from [20] which we call a "false positive rate" and we define its
dual which we call a "false negative rate".

Definition 4. The false positive rate of a trust decision rule f3 associates to

each {G,Gv,s) for s ^ Gy C G the proportion of keys that accept the invalid
statement s:
number of keys k such thai l3{G, k, s) = t r u e
number of keys k in G
Definition 5. The false negative rate of a trust decision rule (i associates to
each (G,Gv,s) for s 6 G„ C G the proportion of keys that reject the valid
statement s:
number of keys k such that l3{G,k,s) = f a l s e
number of keys k in G
In the example of Figures 2 and 3, the trust decision rule has no false posi-
tives. On the other hand, the vahd statement Q: Ignacio is rejected by all keys;
the trust decision rule thus has a false negative rate of 1 for that statement.
Positive results in this direction attempt to establish "stability" of trust
decision rules [20] under degradation or attacks. Figure 4 illustrates the effect of
attacking a trust graph by modification of a single direct trust edge from H -^
P:Nisha to H -^ R: William.

5.4 Valued Graph Models

Thus far, we have considered simple graph models in which edges are unlabeled.
Some models [34, 15] refine these simple graph models by placing constraints on
the trust relationships represented by edges. For instance, E may trust F to issue
 15

Accept P:Nisha
B

Fig. 4. Attack on a trust graph

B [0. 3,3] C
[i.o,2]y<V ^^[0.7,4]
io.s.i.sfv
U).5,2j ^ * P:Nisha
\ . [0.5,1.5]/^
[1.0,1.5]^^/''^ ^*^»--^[0.6,2]
H
• v [0.6,1 / . - * * Q:Ignacio
[0.3,3rV ' y{^A
F - ^ ^ ^ R:Wimam

Fig. 5. Valued tiust graph

valid certificates but not to make good recommendations. Some models assign
values to trust lelationsiiips based on vaiious presumed properties of trust such
as likelihood of failure (of trusting behavior), cost of failure, and an aggregation
from assessments of third parties. The notions of trust state, trust decision rule,
trust attack, and false positive rate apply to these valued graph models as well
and we consider them below.

Network Flow Models. Figure 5 contains a valued trust graph in which each
edge is labeled by a pair of numbers that represents the flow capacity and the
cost of the edge. Trust decision rules include rules where fi{G, k, s) = t r u e only
when the cost per unit flow from k to s is smaller than some threshold, or
when the maximum network flow from k to s exceeds some threshold [20]. Note
that with unit costs and capacities, these metrics reduce to shortest path and
edge-disjoint path metrics for simple trust graphs.

Probabilistic Models. Some categories of trust (see Section 2) are based on

expectations that certain events will occur. For these kinds of trust, P trusts Q
means P expects occurrences of certain events under the control (at least par-
 16

tially) of principal Q. For purposes of discussion let us refer to these events as

trusted events. These events consist of a set of possible outcomes. For instance,
see [3] where trusted events consist of those outcomes in which Q reliably pro-
vides some service.
Accordingly, it is natural to construct trust models in which values associated
to trust relationships are probabilities of events [22, 31]. In [3], [6] the value
associated to a trust relationship is determined by the number of positive trial
outcomes relative to the overall number of trials, while in [17] it is a second order
probability estimate.

Economic Models. Measurement of trust can be regarded as obtaining quan-

titative bounds on risk. This is based on the view that higher levels of trust
correspond to lower levels of risk. Insurance is usually regarded as a means of
transfering financial risk among principals. For instance, the model described in
the paper [29] assigns monetary values to recommendation relationships. The
trust assigned to an assertion is the guaranteed insurance award if the asser-
tion proves to be false. The insurance approach to trust valuation is a promising
direction especially in regard to trust required for commercial electronic transac-
tions where losses attributed to authentication failures can be assessed. However,
this approach only addresses reparable losses (e.g., financial losses); it does not
address irreparable losses such as damage to personal relationships or reputation.

6 Conclusion

Trust plays a critical role in several aspects of agent societies. First, cooperation
between agents depends on their trust in the underlying computer systems and
software. Second, cooperative transactions can incur substantial costs due to ac-
cess barriers between agents; trust between agents enables them to lower these
access barriers thus reducing transaction costs. Finally, trust can lead to the
clustering of agents into communities within which complex multi-agent trans-
actions have a high likelihood of success.
Trust is an overloaded term with numerous meanings, each with distinct
properties. It can be established directly between two or more agents in a variety
of ways. It can be propagated among a network of agents, and one form of trust
can be substituted by other forms. Finally, it can be degraded in several ways.
This paper illustrates the complex nature of trust and the numerous problems
that remain to be solved before software agents can employ computational trust
models.

Acknowledgements. We thank Joshua Guttman, Marion Michaud, and John

Vasak for comments which led to improvements in the presentation of this pa-
per. This work was supported by MITRE's Information Assurance Technical
Operations Program.
 17

References

[1] M. Abadi, M. Burrows, B. Lampson, and G, Plotkin. A calculus for access con-
trol in distributed systems. ACM Transactions on Programnnng Languages and
Systems, 15(4):706-734, October 1993.
[2] S. Berkovits, J. D. Guttman, and V. Swarup. Authentication for mobile agents.
Lecture Notes in Computer Science 1419, Special issue on Mobile Agents and
Security, 1998.
[3] T. Beth, M. Borcherding, and B. Klein. Valuation of trust in open networks.
In D. GoUman, editor. Proceedings of the European Symposium on Research in
Computer Security (ESORICS), LNCS 875, pages 3-18. Springer Verlag, 1994.
[4] A. Birrell, B. Lampson, R. Needham, and M. Shroeder. A global authentication
service without global trust. In Proceedings of the IEEE Symposium on Security
and Privacy, pages 223-230, 1986.
[5] M. Blaze, J. Feigenbaum, and J. Lacy. Decentralized trust management. In
Proceedings of the IEEE Symposium on Security and Privacy, pages 164-173,
1996,
[6] B. Borcherding and M. Borcherding. Covered trust values in distributed systems.
In Proceedings of the Working Conference on Multimedia and Communication
Security, pages 24-31. Chapman & Hall, 1995.
[7] Michael Burrows, Martin Abadi, and Roger Needham. A logic of authentication.
Proceedings of the Royal Society, Series A, 426(1871):233-271, December 1989.
Also appeared as SRC Research Report 39 and, in a shortened form, in ACM
Transactions on Computer Systems 8, 1 (February 1990), 18-36.
[8] G. Edjlali, A. Acharya, and V, Chaudhary. History-based access control for mobile
code. In Proceedings of the ACM Conference on Computer and Communications
Security, 1998.
[9] C. M. Ellison, B. Frantz, B. Lampson, R. R.ivest, B. M. Thomas, and T. Ylonen.
Simple public key certificate. Internet Draft (Work in Progress), November 1998.
[10] C. M. Ellison, B. Frantz, B. Lampson, R,. Rivest, B. M. Thomas, and T. Ylonen.
SPKI certificate theory. Internet Draft (Work in Progress), November 1998.
[11] D. Farmer and W. Venema. SATAN Overview, 1995. http://www.fish.com/.
[12] W. M. Farmer, J. D. Guttman, and V. Swarup. Security for mobile agents:
Authentication and state appraisal. In Proceedings of the European Symposium
on Research in Computer Security (ESORICS), LNCS 1146, pages 118-130, 1996.
[13] F. Fukuyama. TYust : The Social Virtues and the Creation of Prosperity. Free
Press, June 1996.
[14] S. Garfinkel. PGP: Pretty Good Privacy. O'Reilly and Associates, 1994.
[15] E. Gerck. Towards a real-world model of trust: reliance on received information.
MCG, 1998. http://www.mcg.org.br/trustdef.htm.
[16] J. D. Guttman. Filtering postures: Local enforcement for global pohcies. In
Proceedings of the IEEE Sym,posium on Security and Privacy, 1997.
[17] A. Josang. A model for trust in security systems. In Proceedings of the Second
Nordic Workshop on Secure Computer Systems, 1997.
[18] R. M. Kramer and T. R. Tyler, editors. Trust in Organizations : Frontiers of
Theory and Research. Sage Publications, February 1996.
[19] B. Lampson, M. Abadi, M. Burrows, and E. Wobber. Authentication in dis-
tributed systems: Theory and practice. ACM Transactions on Computer Systems,
10(4):265-310, November 1992.
 18

[20] R. Levien and A. Aiken. Attack-resistant trust metrics for public key certification.
In Proceedings of the 7th USENIX Security Symposium, 1998.
[21] S.P. Marsh. FormoMsing Trust as a Computational Concept. PhD thesis, De-
partment of Computer Science and Mathematics, University of Sterhng, April
1994.
[22] U. Maurer. Modeling a public-key infrastructure. In Proceedings of the European
Symposium on Research in Computer Security (ESORICS), LNCS 1146, pages
118-130. Springer Verlag, 1996.
[23] D.H. McKnight and N.L. Chervany. The meanings of trust. Working pa-
per, Carlson School of Management, University of Minnesota, 1996. h t t p : / /
www. misrc. uiim. edu/wpaper/wp96-04. htm.
[24] B. A. Misztal. Trust m Modern Societies : The Search for the Bases of Social
Order. Polity Press, December 1995.
[25] P. Venkat Rangan. An axiomatic bsisis of trust in distributed systems. In Pro-
ceedings of the IEEE Symposium on Security and Privacy, pages 204-210, 1988.
[26] J.M. Reagle. Trust in a cryptographic economy and digital security deposits:
Protocols and policies. Master's thesis. Technology and Policj' Program, Mas-
sachusetts Institute of Technology, May 1996.
[27] J.M. Reagle. Trust in electronic markets: The convergence of cryptographers and
economists. First Monday, 1(2), August 1996. http://www.firstmonday.dk/
issues/issue2/markets/index.html.
[28] M. K. Reiter and S. G. Stubblebine. Path independence for authentication in
large-scale systems. In Proceedings of the 4th ACM Conference on Computer and
Communications Security, pages 57-66, 1997.
[29] M. K. Reiter and S. G. Stubblebine. Toward acceptable metrics of authentication.
In Proceedings of the IEEE Symposium on Security and Privacy, pages 3-18, 1997.
[30] T. Sander and C. Tschudin. Towards mobile cryptography. In Proceedings of the
IEEE Symposium on Security and Privacy, 1998.
[31] A. Tarah and C. Huitema. Associating metrics to certification paths. In Proceed-
ings of the European Symposivjn on Research in Computer Security (ESORICS),
LNCS 648, pages 175-189. Springer Verlag, 1992.
[32] Bernard Wilhams. Formal structures and social reality. In D. Gambetta, editor,
Trust: Making and Breaking Cooperative Relations, pages 3-13. Basil Blackwell,
1988.
[33] R. Yahalom, B. Klein, and Th. Beth. Trust relationships in secure systems - a
distributed authentication perspective. In Proceedings of the IEEE Symposium
on Security and Privacy, 1993.
[34] R. Yahalom, B. Klein, and Th. Beth. Trust-based navigation in distributed sys-
tems. Computing Systems, 7(l):45-73, 1994.
 Protection in
Programming-Language Translations*

Martin Abadi

A b s t r a c t . We discuss abstractions for protection and the correctness

of their implementations. Relying on the concept of full abstraction, we
consider two examples: (1) the translation of Java classes to an inter-
mediate bytecode language, and (2) in the setting of the pi calculus, the
implementation of private channels in terms of cryptographic operations.

1 Introduction
Tangible crimes and measures against those crimes are sometimes explained
through abstract models—with mixed results, as the detective Erik Lonnrot
discovered [9]. Protection in computer systems relies on abstractions too. For
example, an access matrix is a high-level specification t h a t describes the allowed
accesses of subjects to objects in a computer system; the system may rely on
mechanisms such as access lists and capabilities for implementing an access ma-
trix [22].
Abstractions are often embodied in programming-language constructs. Re-
cent work on Java [17] has popularized the idea t h a t languages are relevant
t o security, but the relation between languages and security is much older. In
particular, objects and types have long been used for protection against incompe-
tence and malice, a t least since the 1970s [36, 24, 20]. In the realm of distributed
systems, programming languages (or their libraries) have sometimes provided ab-
stractions for communication on secure channels of the kind implemented with
cryptography [7,49,47,50,46].
Security depends not only on the design of clear and expressive abstractions
but also on the correctness of their implementations. Unfortunately, the crite-
ria for correctness are rarely stated precisely—and presumably they are rarely
met. These criteria seem particularly delicate when a principal relies on those
abstractions but interacts with other principals at a lower level. For example,
the principal may express its programs and policies in terms of objects and re-
mote method invocations, but may send and receive bit strings. Moreover, t h e
bit strings t h a t it receives may not have been the o u t p u t of software t r u s t e d to
respect the abstractions. Such situations seem to be more common now t h a n in
the 1970s.
One of the difficulties in the correct implementation of secure systems is
t h a t the s t a n d a r d notion of refinement (e.g., [19,21]) does not preserve security

* This is a slightly revised version of a paper that appeared in Automata, Lan-

guages and Programming: 25th International Colloquium, ICALP'98, Springer-
Verlag, July 1998.
 20

properties. Ordinarily, the non-determinism of a specification may be intended

to allow a variety of implementations. In security, the non-determinism may also
serve for hiding sensitive data. As an example, let us consider a specification that
describes a computer that displays an arbitrary but fixed string in a corner of
a screen. A proposed implementation might always display a user's password as
that string. Although this implementation may be functionally correct, we may
consider it incorrect for security purposes, because it leaks more information
than the specification seems to allow. Security properties are thus different from
other common properties; in fact, it has been argued that security properties do
not conform to the usual definition of properties as predicates on behaviors [6,
21,28].
Reexamining this example, let us write P for the user's password, I{P) for
the proposed implementation, and S{P) for the specification. Since the set of
behaviors allowed by the specification does not depend on P, clearly S{P) is
equivalent to S{P') for any other password P'. On the other hand, I{P) and
I{P') are not equivalent, since an observer can distinguish them. Since the map-
ping from specification to implementation does not preserve equivalence, we may
say that it is not fully abstract [30,40]. We may explain the perceived weakness
of the proposed implementation by this failure of full abstraction.
This paper suggests that, more generally, the concept of full abstraction is a
useful tool for understanding the problem of implementing secure systems. Full
abstraction seems particularly pertinent in systems that rely on translations
between languages—for example, higher-level languages with objects and secure
channels, lower-level languages with memory addresses and cryptographic keys.
We consider two examples of rather different natures and review some stan-
dard security concerns, relating these concerns to the pursuit of full abstraction.
The first example arises in the context of Java (section 2). The second one con-
cerns the implementation of secure channels, and relies on the pi calculus as
formal framework (section 3). The thesis of this paper about full abstraction is
in part a device for discussing these two examples.
This paper is rather informal and partly tutorial; its contributions are a
perspective on some security problems and some examples, not new theorems.
Related results appear in more technical papers [44,2].

Full abstraction, revisited

We say that two expressions are equivalent in a given language if they yield
the same observable results in all contexts of the language. A translation from
a language Li to a language L2 is equationally fully abstract if (1) it maps
equivalent Li expressions to equivalent L2 expressions, and (2) conversely, it
maps nonequivalent Li expressions to nonequivalent L2 expressions [30,40,43,
35,42]. We may think of the context of an expression as an attacker that interacts
with the expression, perhaps trying to learn some sensitive information (e.g., [3]).
With this view, condition (1) means that the translation does not introduce
information leaks. Since equations may express not only secrecy properties but
also some integrity properties, the translation must preserve those properties as
 21

well. Because of these consequences of condition (1), we focus on it; we mostly

ignore condition (2), although it can be useful too, in particular for excluding
trivial translations.
Closely related to equational full abstraction is logical full abstraction [27].
A translation from a language Li to a language L2 is logically fully abstract
if it preserves logical properties of the expressions being translated. Longley
and Plotkin have identified conditions under which equational and logical full
abstraction are equivalent. Since we use the concept of full abstraction loosely,
we do not distinguish its nuances.
An expression of the source language L\ may be written in a silly, incompe-
tent, or even malicious way. For example, the expression may be a program that
broadcasts some sensitive information—so this expression is insecure on its own,
even before any translation to L-z- Thus, full abstraction is clearly not sufficient
for security; however, as we discuss in this paper, it is often relevant.

2 Objects and Mobile Code

The Java programming language is typically compiled to an intermediate lan-
guage, which we call JVML and which is implemented by the Java Virtual Ma-
chine [17,26]. JVML programs are often communicated across networks, for ex-
ample from Web servers to their chents. A chent may run a JVML program in
a Java Virtual Machine embedded in a Web browser. The Java Virtual Machine
helps protect local resources from mobile JVML programs while allowing those
programs to interact with local class libraries. Some of these local class libraries
perform essential functions (for example, input and output), so they are often
viewed as part of the Java Virtual Machine.

2.1 Translating Java to JVML

As a first example we consider the following trivial Java class:
class C {
private int x;
public void set_x(int v) {
this.x = v;
};
}
This class describes objects with a field x and a method set_x. The method
set_x takes an integer argument v and updates the field x to v. The keyword
this represents the self of an object; the keyword public indicates that any client
or subclass can access set_x directly; the keyword private disallows a similar
direct access to x from outside the class. Therefore, the field x can be written
but never read.
The result of compiling this class to JVML may be expressed roughly as
follows. (Here we do not use the official, concrete syntax of JVML, which is not
designed for human understanding.)
 22

sC {
private int x;
public void set_x(int) {
.framelimits l o c a l s = 2, s tack 2;
aload_0; II load this
iload-1; II load V
p u t f i e l d x; II s e t X
};
}

As this example indicates, JVML is a fairly high-level language, and in particular

it features object-oriented constructs such as classes, methods, and self. It differs
from Java in that methods manipulate local variables, a stack, and a heap using
low-level load and store operations. The details of those operations are not im-
portant for our purposes. Each method body declares how many local variables
and stack slots its activation may require. The Java Virtual Machine includes
a bytecode verifier, which checks that those declarations are conservative (for
instance, that the stack will not overflow). If undetected, dynamic errors such
as stack overflow could lead to unpredictable behavior and to security breaches.
The writer of a Java program may have some security-related expectations
about the program. In our simple example, the field x cannot be read from
outside the class, so it may be used for storing sensitive information. Our example
is so trivial that this information cannot be exploited in any way, but there are
more substantial and interesting examples that permit controlled access to fields
with the qualifier private and similar qualifiers. For instance, a Java class for
random-number generation (like java.util.Random) may store seeds in private
fields. In these examples, a security property of a Java class may be deduced—
or presumed—by considering all possible Java contexts in which the class can
be used. Because those contexts must obey the type rules of Java, they cannot
access private fields of the class.
When a Java class is translated to JVML, one would hke the resulting JVML
code to have the security properties that were expected at the Java level. How-
ever, the JVML code interacts with a JVML context, not with a Java context.
If the translation from Java to JVML is fully abstract, then matters are con-
siderably simphfied—in that case, JVML contexts have no more power than
Java contexts. Unfortunately, as we point out below, the current translation is
not fully abstract (at least not in a straightforward sense). Nevertheless, the
translation approximates full abstraction:

— In our example, the translation retains the qualifier private for x. The oc-
currence of this qualifier at the JVML level may not be surprising, but it
cannot be taken for granted. (At the JVML level, the qualifier does not have
the benefit of helping programmers adhere to sound software-engineering
practices, since programmers hardly ever write JVML, so the qualifier might
have been omitted from JVML.)
 23

- Furthermore, the bytecode verifier can perform standard typechecking, guar-

anteeing in particular that a JVML class does not refer to a private field of
another JVML class.
- The bytecode verifier can also check that dynamic errors such as stack over-
flow will not occur. Therefore, the behavior of JVML classes should conform
to the intended JVML semantics; JVML code cannot get around the JVML
type system for accessing a private field inappropriately.
Thus, the bytecode verifier restricts the set of JVML contexts, and in effect
makes them resemble Java contexts (cf. [17, p. 220]). As the set of JVML contexts
decreases, the set of equivalences satisfied by JVML programs increases, so the
translation from Java to JVML gets closer to full abstraction. Therefore, we
might even view full abstraction as the goal of bytecode verification.
Recently, there have been several rigorous studies of the Java Virtual Ma-
chine, and in particular of the bytecode verifier [10,44,41,16]. These studies
focus on the type-safety of the JVML programs accepted by the bytecode veri-
fier. As has long been believed, and as Leroy and Rouaix have recently proved in
a somewhat different context [25], strong typing yields some basic but important
security guarantees. However, those guarantees do not concern language trans-
lations. By themselves, those guarantees do not imply that libraries written in
a high-level language have expected security properties when they interact with
lower-level mobile code.

2.2 Obstacles to full abstraction

As noted, the current translation of Java to JVML is not fully abstract. The
following variant of our first example illustrates the failure of full abstraction.
We have no reason to believe that it illustrates the only reason for the failure of
full abstraction, or the most worrisome one; Dean, Felten, Wallach, and Balfanz
have discovered several significant discrepancies between the semantics of Java
and that of JVML [12].
class D {
class E {
private int y = x;
};_
private int x;
public void set_x(int v) {
this.x = v;
);
}
The class E is an inner class [45]. To each instance of an inner class such as E
corresponds an instance of its outer class, D in this example. The inner class may
legally refer to the private fields of the outer class.
Unlike Java, JVML does not include an inner-class construct. Therefore,
compilers "flatten" inner classes while adding accessor methods. Basically, as far
 24

as compilation is concerned, we may as well have written the following classes

instead of D:

class D {
private int x;
public void set_x(int v) {
t h i s . x = v;
};
static int get_x(D d) {
r e t u r n d.x;
};
}

class E {
. . . get-jc . ..
}
Here E is moved to the top level. A method get_x is added to D and used in E
for reading x; the details of E do not matter for our purposes. The method get_x
can be used not just in E, however—any other class within the same package
may refer to get jx.
When the classes D and E are compiled to JVML, therefore, a JVML context
may be able to read x in a way that was not possible at the Java level. This pos-
sibihty results in the loss of full abstraction, since there is a JVML context that
distinguishes objects that could not be distinguished by any Java context. More
precisely, a JVML context that runs get_x and returns the result distinguishes
instances of D with different values for x.
This loss of full abstraction may result in the leak of some sensitive infor-
mation, if any was stored in the field x. The leak of the contents of a private
component of an object can be a concern when the object is part of the Java
Virtual Machine, or when it is trusted by the Java Virtual Machine (for exam-
ple, because a trusted principal digitally signed the object's class). On the other
hand, when the object is part of an applet, this leak should not be surprising:
applets cannot usually be protected from their execution environments.
For better or for worse, the Java security story is more complicated and
dynamic than the discussion above might suggest. In addition to protection by
the qualifier private, Java has a default mode of protection that protects classes
in one package against classes in other packages. At the language level, this mode
of protection is void—any class can claim to belong to any package. However,
Java class loaders can treat certain packages in special ways, guaranteeing that
only trusted classes belong to them. Our example with inner classes does not
pose a security problem as long as D and E are in one of those packages.
In hindsight, it is not clear whether one should base any security expectations
on qualifiers like private, and more generally on other Java constructs. As Dean
et al. have argued [12], the definition of Java is weaker than it should be from a
security viewpoint. Although it would be prudent to strengthen that definition,
 25

a full-blown requirement of full abstraction may not be a necessary addition.

More modest additions may suffice. Section 4 discusses this subject further.

3 Channels for Distributed Communication

In this section, we consider the problem of implementing secure channels in

distributed systems. As mentioned in the introduction, some systems for dis-
tributed programming offer abstractions for creating and using secure channels.
The implementations of those channels typically rely on cryptography for en-
suring the privacy and the integrity of network communication. The relation
between the abstractions and their implementations is usually explained only
informally. Moreover, the abstractions are seldom explained in a self-contained
manner that would permit reasoning about them without considering their im-
plementations at least occasionally.
The concept of full abstraction can serve as a guide in understanding secure
channels. When trying to approximate full abstraction, we rediscover common
attacks and countermeasures. Most importantly, the pursuit of full abstraction
entails a healthy attention to the connections between an implementation and
higher-level programs that use the implementation, beyond the intrinsic proper-
ties of the implementation.

3.1 Translating the pi calculus to the spi calculus

The formal setting for this section is the pi calculus [32,34,33], which serves
as a core calculus with primitives for creating and using channels. By applying
the pi calculus restriction operator, these channels can be made private. We dis-
cuss the problem of mapping the pi calculus to a lower-level calculus, the spi
calculus [4,5,3], implementing communication on private channels by encrypted
communication on public channels. Several low-level attacks can be cast as coun-
terexamples to the full abstraction of this mapping. Some of the attacks can be
thwarted through techniques common in the literature on protocol design [29].
Some other attacks suggest fundamental difficulties in achieving full abstraction
for the pi calculus.
First we briefly review the spi calculus. In the variant that we consider here,
the syntax of this calculus assumes an infinite set of names and an infinite set
of variables. We let c, d, m, n, and p range over names, and let w, x, y, and z
range over variables. We usually assume that all these names and variables are
different (for example, that m and n are different names). The set of terms of
the spi calculus is defined by the following grammar:

L,M,N::= terms
n name
X variable
{ M l , . . . ,M(fc}jv encryption (A; > 0)
 26

Intuitively, { M i , . . . , Mfc}jv represents the ciphertext obtained by encrypting the

terms M i , . . . , Mk under the key N (using a symmetric cryptosystem such as
DES or RC5 [29]). The set of processes of the spi calculus is defined by the
following grammar:

P,Q ::= processes

'M{Ni,...,Nk) output (A; > 0 )
M{xi,...,Xk).P input (A; > 0)
0 nil
P IQ composition
IP replication
{vn)P restriction
[M is N] P match
case L of {xi,... ,Xk}N in P decryption {k > 0)

An output process M{Ni,... ,Nk) sends the tuple Ni,... ,Nk on M. An input
process M(a:i,... ,Xk)-Q is ready to input k terms Ni,... ,Nk on M, and then
to behave as Q[N\/xi,... ,Ni./xk\- Here we write Q[Ni/xi,... ,Nk/xk] for the
result of replacing each free occurrence of xt in Q with Ni, for i 6 l..k. Both
M{xi,... ,Xk)-Q and case L of {xi,... ,Xk]}\/ in P (explained below) bind the
variables xi, ..., x^. The nil process 0 does nothing. A composition P \ Q
behaves as P and Q running in parallel. A replication !P behaves as infinitely
many copies of P running in parallel. A restriction (i/n)P makes a new name
n and then behaves as P; it binds the name n. A match process [M is N] P
behaves as P if M and TV are equal; otherwise it does nothing. A decryption
process case L of [xi,... ,Xk}N in P attempts to decrypt L with the key TV; if L
has the form { M i , . . . , M^JN, then the process behaves as P [ M i / 2 : i , . . . , Mk/xk];
otherwise it does nothing.
By omitting the constructs { M i , . . . ,Mfc}jv and case L of {xi,... ,Xk}N in P
from these grammars, we obtain the syntax of the pi calculus (more precisely, of
a polyadic, asynchronous version of the pi calculus).
As a first example, we consider the trivial pi calculus process:

{i/n){n{m) \ n(x).O)

This is a process that creates a channel n, then uses it for transmitting the name
m, with no further consequence. Communication on n is secure in the sense that
no context can discover m by interacting with this process, and no context can
cause a different message to be sent on n; these are typical secrecy and integrity
properties. Such properties can be expressed as equivalences (in particular, as
testing equivalences [11,8,3]). For example, we may express the secrecy of m as
the equivalence between {i'n){n{m) \ n(x).O) and {i'n){n(m') \ n{x).0), for any
names m and m'.
Intuitively, the subprocesses n(m) and n{x).0 may execute on different ma-
chines; the network between these machines may not be physically secure. There-
fore, we would like to explicate a channel like n in lower-level terms, mapping it
 27

to some sort of encrypted connection multiplexed on a public channel. For exam-

ple, we might translate our first process, (i/n)(n(m) | n{x).0), into the following
spi calculus process:

(!/n)(c({m}„) I c{y).case y of {a;}„ in 0)

Here c is a distinguished, free name, intuitively the name of a well-known public

channel. The name n still appears, with a restriction, but it is used for a key
rather than for a channel. The sender encrypts m using n; the recipient tries to
decrypt a ciphertext y that it receives on c using n; if the decryption succeeds,
the recipient obtains a cleartext x (hopefully m).
This translation strategy may seem promising. However, it has numerous
weaknesses; we describe several of those weaknesses in what follows. The weak-
nesses represent obstacles to full abstraction and are also significant in practical
terms.

3.2 Obstacles to full abstraction

Leak of traffic patterns In the pi calculus, (i'n){n(m) \ n{x).0) is simply
equivalent to 0, because the internal communication on n cannot be observed.
On the other hand, in the spi calculus, (i/n)(c({m}„) | c{y).case y of {x}n in 0)
is not equivalent to the obvious implementation of 0, namely 0. A spi calculus
process that interacts with (i/n)(c({m}„) | c{y).case y of {x}n in 0) can detect
traffic on c, even if it cannot decrypt that traffic.
The obvious way to protect against this leak is to add noise to communication
lines. In the context of the spi calculus, we may for example compose all our
implementations with the noise process ^•{i'p)c{{}p)- This process continually
generates keys and uses those keys for producing encrypted traffic on the public
channel c.
In practice, since noise is rather wasteful of communication resources, and
since a certain amount of noise might be assumed to exist on communication lines
as a matter of course, noise is not always added in implementations. Without
noise, full abstraction fails.

Trivial denial-of-service vulnerability Consider the pi calculus process

{i'Ti){n(m) I n{x).x{))

which is a small variant of the first example where, after its receipt, the message
m is used for sending an empty message. This process preserves the integrity
of m, in the sense that no other name can be received and used instead of m;
therefore, this process is equivalent to m().
The obvious spi calculus implementations of {i'n){n{Tn) \ n{x).x{)) and m()
are respectively

{vn){c{{m}n) I c{ij).case,y of {a;}„ in c{{]x))

 28

and c({}m). These implementations can be distinguished not only by traffic

analysis but also in other trivial ways. For example, the former implementa-
tion may become stuck when it interacts with c(p), because the decryption
case y of {x}n in c{{}x) fails when y is p rather than a ciphertext. In con-
trast, the latter implementation does not suffer from this problem.
Informally, we may say that the process c(p) mounts a denial-of-service at-
tack. Formally, such attacks can sometimes be ignored by focusing on process
equivalences that capture only safety properties, and not liveness properties.
In addition, the implementations may be strengthened, as is commonly done
in practical systems. For example, as a first improvement, we may add some
replication to (i/n){c{{m}n) \ c{y).case y of {a;}„ in c({}x)), obtaining:

(i/n)(c({m}„) I \c{y).case y of {x}n in c{{}x))

This use of replication protects against c(p}.

Exposure to replay attacks Another shortcoming of our implementation

strategy is exposure to replay attacks. As an example, we consider the pi calculus
process:
(i/n)(n(mi) | n(m2) | n{x).x{) \ n{x).x{))
which differs from the previous example only in that two names mi and m2 are
transmitted on n, asynchronously. In the pi calculus, this process is equivalent
to mr() I m2"(): it is guaranteed that both TOI and m2 go from sender to receiver
exactly once.
This guarantee is not shared by the spi calculus implementation

/c({mi}„) I c({m2}n) I \
(i/n) c{y).case y of {a;}„ in c{{}x) \ 1
\c{y).case y of {x}„ in c{{}x) J

independently of any denial-of-service attacks. When this implementation is

combined with the spi calculus process c{y).{c{y) \ c{y)), which duplicates a
message on c, two identical messages may result, either c({}mi) | c{{}mi) or
C({}m2) I C({}TO,).
Informally, we may say that the process c{y).{c{y) \ c{y)) mounts a re-
play attack. Standard countermeasures apply: timestamps, sequence numbers,
and challenge-response protocols. In this example, the addition of a minimal
challenge-response protocol leads to the following spi calculus process:

( c{zi).c{{mi,Zi]n) I c{z2).c{{m2,Z2]n) \

{vpi){c{pi) I c{y).case y of {x,Zi]n in [zi ispi]c{{]x)) \

\

{vp2){c{p2) I c{y).case y of {x,Z2]n in [z2 isp2]c{{}x)) J

The names pi and p2 serve as challenges; they are sent by the subprocesses that
are meant to receive mi and ma, received by the subprocesses that send mi
and m.2, and included along with mi and m2 under n. This challenge-response
 29

protocol is rather simplistic in that the challenges may get "crossed" and then
neither mi nor m2 would be transmitted successfully; it is a simple matter of
programming to protect against this confusion. In any case, for each challenge,
at most one message is accepted under n. This use of challenges thwarts replay
attacks.

Leak of message equalities In the pi calculus, the identity of messages sent

on private channels is concealed. For example, an observer of the process

{vn){n{mi) | n(m2) | n{x).0 \ n{x).0)

will not even discover whether mi —1712. (For this example, we drop the impHcit
assumption that mi and m2 are different names.) On the other hand, suppose
that we translate this process to:

( c({mi}„) I c({m2}„) I
c{y).case y of {a;}„ in 0 |
c{y).case y of {x}n in 0 J
"^

An observer of this process can tell whether mi = m j , even without knowing

mi or m2 (or n). In particular, the observer may execute:

c{x).c{y).{[x is ij] d{) \ c{x) \ c{y))

This process reads and relays two messages on the channel c, and emits a message
on the channel d if the two messages are equal. It therefore distinguishes whether
mi = m2. The importance of this sort of leak depends on circumstances. In an
extreme case, one cleartext may have been guessed (for example, the cleartext
"attack at dawn"); knowing that another message contains the same cleartext
may then be significant.
A simple countermeasure consists in including a different confounder com-
ponent in each encrypted message. In this example, the implementation would
become:
I {'^Pi.)c{{'m-uPl]n) I {l'P2)c{{m2,p2]n) \\
{vn) c{y).case y of {x,zi}n in 0 \
\ciij).case y of {x,Z2}„ in 0 J
The names pi and p2 are used only to differentiate the two messages being
transmitted. Their inclusion in those messages ensures that a comparison on
ciphertexts does not reveal an equality of cleartexts.

Lack of forward secrecy As a final example, we consider the pi calculus

process:
{i'n){n{m) | n{x).p{n))
This process transmits the name m on the channel n, which is private until this
point. Then it releases n by sending it on the public channel p. Other processes
 30

may use n afterwards, but cannot recover the contents of the first message sent
on n. Therefore, this process is equivalent to
(i/n)(n(m') | n{x).p{n))
for any m'. Interestingly, this example relies crucially on scope extrusion, a
feature of the pi calculus not present in simpler calculi such as CCS [31].
A spi calculus implementation of {i'n){n{m) | n{x).p{n)) might be:
{i^n){c{{m]n) I c{y).case y of {x}n in c{{n]p))
However, this implementation lacks the forward-secrecy property [14]: the disclo-
sure of the key n compromises all data previously sent under n. More precisely,
a process may read messages on c and remember them, obtain n by decrypt-
ing {n}p, then use n for decrypting older messages on c. In particular, the spi
calculus process
c{x).{c{x) I c{y).case y of {z}p in case x of {w}^ in d(w))
may read and relay {m}„, read and decrypt {n}p, then go back to obtain m
from {m}„, and finally release m on the public channel d.
Full abstraction is lost, as with the other attacks; in this case, however, it
seems much harder to recover. Several solutions may be considered.
- We may restrict the pi calculus somehow, ruling out troublesome cases of
scope extrusion. It is not immediately clear whether enough expressiveness
for practical programming can be retained.
- We may add some constructs to the pi calculus, for example a construct that
given the name n of a channel will yield all previous messages sent on the
channel n. The addition of this construct will destroy the source-language
equivalence that was not preserved by the translation. On the other hand,
this construct seems fairly artificial.
- We may somehow indicate that source-language equivalences should not be
taken too seriously. In particular, we may reveal some aspects of the imple-
mentation, warning that forward secrecy may not hold. We may also spec-
ify which source-language properties are maintained in the implementation.
This solution is perhaps the most realistic one, although we do not yet know
how to write the necessary specifications in a precise and manageable form.
- Finally, we may try to strengthen the implementation. For example, we may
vary the key that corresponds to a pi calculus channel by, at each instant,
computing a new key by hashing the previous one. This approach is fairly
elaborate and expensive.
The problem of forward secrecy may be neatly avoided by shifting from the pi
calculus to the join calculus [15]. The join calculus separates the capabilities for
sending and receiving on a channel, and forbids the communication of the latter
capability. Because of this asymmetry, the join calculus is somewhat easier to
map to a lower-level calculus with cryptographic constructs. This mapping is the
subject of current work [2]; although still impractical, the translation obtained
is fully abstract.
 31

4 Full Abstraction in Context

With progress on security infrastructures and techniques, it may become less im-
portant for translations to approximate full abstraction. Instead, we may rely on
the intrinsic security properties of target-language code and on digital signatures
on this code. We may also rely on the security properties of source-language
code, but only when a precise specification asserts that translation preserves
those properties. Unfortunately, several caveats apply.

- The intrinsic security properties of target-language code may be extremely

hard to discover a posteriori. Languages such as JVML are not designed for
ease of reading. Furthermore, the proof of those properties may require the
analysis of delicate and complex cryptographic protocols. Certifying compil-
ers [39,37] may alleviate these problems but may not fully solve them.
- Digital signatures complement static analyses but do not obviate them. In
particular, digital signatures cannot protect against incompetence or against
misplaced trust. Moreover, digital signatures do not seem applicable in all
settings. For example, digital signatures on spi calculus processes would be of
little use, since these processes never migrate from one machine to another.
- Finally, we still have only a limited understanding of how to specify and
prove that a translation preserves particular security properties. This ques-
tion deserves further attention. It may be worthwhile to address it first in
special cases, for example for information-flow properties [13] as captured in
type systems [48,1,38,18].

The judicious use of abstractions can contribute to simplicity, and thus

to security. On the other hand, abstractions and their translations can give
rise to complications, subtleties, and ultimately to security flaws. As Lampson
wrote [23], "neither abstraction nor simplicity is a substitute for getting it right".
Concepts such as full abstraction should help in getting it right.

Acknowledgements

Most of the observations of this paper were made during joint work with Cedric
Fournet, Georges Gonthier, Andy Gordon, and Raymie Stata. Drew Dean, Mark
Lillibridge, and Dan Wallach helped by explaining various Java subtleties. Mike
Burrows, Cedric Fournet, Mark Lillibridge, John Mitchell, and Dan Wallach
suggested improvements to a draft. The title is derived from that of a paper by
Jim Morris [36].

References

1. Martin Abadi. Secrecy by typing in security protocols. In Theoretical Aspects

of Computer Software, volume 1281 of Lecture Notes in Computer Science, pages
611-638. Springer-Verlag, 1997.
 32

2. Martin Abadi, Cedric Fournet, and Georges Gonthier. Secure implementation of

channel abstractions. In Proceedings of the Thirteenth Annual IEEE Symposium
on Logic in Computer Science, pages 105-116, June 1998.
3. Martin Abadi and Andrew D. Gordon. A calculus for cryptographic protocols: The
spi calculus. TechnicaJ Report 414, University of Cambridge Computer Laboratory,
January 1997. Extended version of both [4] and [5]. A revised version appeared as
Digital Equipment Corporation Systems Research Center report No. 149, January
1998, and an abridged version will appear in Information and Computation.
4. Martin Abadi and Andrew D. Gordon. A calculus for cryptographic protocols:
The spi calculus. In Proceedings of the Fourth ACM Conference on Computer and
Communications Security, pages 36-47, 1997.
5. Martin Abadi and Andrew D. Gordon. Reasoning about cryptographic protocols in
the spi calculus. In Proceedings of the 8th International Conference on Concurrency
Theory, volume 1243 of Lecture Notes in Computer Science, pages 59-73. Springer-
Verlag, July 1997.
6. Bowen Alpern and Fred B. Schneider. Defining liveness. Information Processing
Letters, 21(4):181-185, October 1985.
7. Andrew D. Birrell. Secure communication using remote procedure calls. ACM
Transactions on Computer Systems, 3(1):1-14, February 1985.
8. Michele Boreale and Rocco De Nicola. Testing equivalence for mobile processes.
Information and Computation, 120(2):279-303, August 1995.
9. Jorge Luis Borges. La muerte y la briijula. In Obras completas 1923-1972, pages
499-507. Emece Editores, Buenos Aires, 1974. Titled "Death and the compass" in
English translations.
10. Richard M. Cohen. Defensive Java Virtual Machine version 0.5 alpha release. Web
pages at http://www.cli.com/. May 1997.
11. Rocco De Nicola and Matthew C. B. Hennessy. Testing equivalences for processes.
Theoretical Computer Science, 34:83-133, 1984.
12. Drew Dean, Edwcird W. Felten, Dan S. Wallach, and Dirk Balfanz. Java secu-
rity: Web browsers and beyond. In Dorothy E. Denning and Peter J. Denning,
editors, Internet besieged: countering cyberspace scofflaws, pages 241-269. ACM
Press, 1998.
13. Dorothy E. Denning. Cryptography and Data Security. Addison-Wesley, Reading,
Mass., 1982.
14. Whitfield Diflie, Paul C. van Oorschot, and Michael J. Wiener. Authentication
and authenticated key exchanges. Designs, Codes and Cryptography, 2:107-125,
1992.
15. Cedric Fournet and Georges Gonthier. The reflexive chemical abstract machine
and the join-calculus. In Proceedings of the 23rd ACM Symposium on Principles
of Programming Languages, pages 372-385, January 1996.
16. Stephen N. Freund and John C. Mitchell. A type system for object initializa-
tion in the Java bytecode language. In OOPSLA '98 Conference Proceedings:
Object-Oriented Programming, Systems, Languages, and Applications, pages 310-
327, 1998.
17. James Gosling, Bill Joy, and Guy L. Steele. The Java Language Specification.
Addison-Wesley, 1996.
18. Nevin Heintze and Jon G. Riecke. The SLam calculus: programming with se-
crecy and integrity. In Proceedings of the 25th ACM Symposium on Principles of
Programming Languages, pages 365-377, 1998.
19. C. A. R. Hoare. Proof of correctness of data representations. Acta Informatica,
1:271-281, 1972.
 33

20. Anita K. Jones and Barbara H. Liskov. A language extension for expressing con-
straints on data access. Communications of the ACM, 21(5):358-367, May 1978.
21. Leslie Lamport. A simple approach to specifying concurrent systems. Com.m.uni-
cations of the ACM, 32(l):32-45, January 1989.
22. Butler W. Lampson. Protection. In Proceedings of the 5th Princeton Conference
on Information Sciences and Systems, pages 437-443, 1971.
23. Butler W. Lampson. Hints for computer system design. Operating Systems Re-
view, 17(5):33-48, October 1983. Proceedings of the Ninth ACM Symposium on
Operating System Principles.
24. Butler W. Lampson and Howard E. Sturgis. Reflections on an operating system
design. Communications of the ACM, 19(5):251-265, May 1976.
25. Xavier Leroy and Frangois Rouaix. Security properties of typed applets. In Pro-
ceedings of the 25th ACM Symposium on Principles of Programming Languages,
pages 391-403, 1998.
26. Tim Lindholm and Frank Yellin. The Java Virtual Machine Specification. Addison-
Wesley, 1996.
27. John Longley and Gordon Plotkin. Logical full abstraction and PCF. In Jonathan
Ginzburg, Zurab Khasidashvili, Carl Vogel, Jean-Jacques Levy, and Enric Vallduvi,
editors, The Tbilisi Symposium on Logic, Language and Computation: Selected
Papers, pages 333-352. CSLI Pubhcations and FoLLI, 1998.
28. John McLean. A general theory of composition for a class of "possibilistic" prop-
erties. IEEE Transactions on Software Engineering, 22(l):53-66, January 1996.
29. Alfred J. Menezes, Paul C. van Oorschot, and Scott A. Vanstone. Handbook of
Applied Cryptography. CRC Press, 1996.
30. Robin Milner. Fully abstract models of typed A-calculi. Theoretical Computer
Science, 4:1-22, 1977.
31. Robin Milner. Communication and Concurrency. Prentice-Hall International,
1989.
32. Robin Milner. Functions as processes. Mathematical Structures in Computer Sci-
ence, 2:119-141, 1992.
33. Robin Milner. The polyadic 7r-calculus: a tutorial. In Bauer, Brauer, and Schwicht-
enberg, editors, Logic and Algebra of Specification. Springer-Verlag, 1993.
34. Robin Milner, Joachim Parrow, and David Walker. A calculus of mobile processes,
parts I and II. Information and Computation, 100:1-40 and 41-77, September
1992.
35. John C. Mitchell. On abstraction and the expressive power of programming lan-
guages. Science of Computer Programming, 21(2):141-163, October 1993.
36. James H. Morris, Jr. Protection in programming languages. Communications of
the ACM, 16(1):15-21, January 1973.
37. Greg Morrisett, David Walker, Karl Crary, and Neal Glew. Prom System F to
Typed Assembly Language. In Proceedings of the 25th ACM Symposium on Prin-
ciples of Programming Languages, pages 85-97, 1998.
38. Andrew C. Myers and Barbara Liskov. A decentralized model for information
flow control. In Proceedings of the 16th ACM Symposium on Operating System
Principles, pages 129-142, 1997.
39. George C. Necula and Peter Lee. The design and implementation of a certifying
compiler. In Proceedings of the ACM SIGPLAN'98 Conference on Programming
Language Design and Implementation (PLDI), pages 333-344, 1998.
40. Gordon Plotkin. LCF considered as a programming language. Theoretical Com-
puter Science, 5:223-256, 1977.
 34

41. Zhenyu Qian. A formal specification of Java Virtual Machine instructions for
objects, methods and subroutines. In Jim Alves-Foss, editor, Formal Syntax and
Semantics of Java'^". Springer-Verlag, 1998. To appear.
42. Jon G. Riecke. Fully abstract translations between functional languages. Mathe-
matical Structures in Computer Science, 3(4):387-415, December 1993.
43. Ehud Shapiro. Separating concurrent languages with categories of language em-
beddings. In Proceedings of the Twenty Third Annual ACM Symposium on the
Theory of Computing, pages 198-208, 1991.
44. Raymie Stata and Martin Abadi. A type system for Java bytecode subroutines. In
Proceedings of the 25th ACM Symposium on Principles of Programming Languages,
pages 149-160, January 1998.
45. Sun Microsystems, Inc. Inner classes specification. Web pages at h t t p : / / j a v a
. s u n . c o m / p r o d u c t s / j d k / 1 . 1 / d o c s / g u i d e / i n n e r c l a s s e s / , 1997.
46. Sun Microsystems, Inc. RMI enhancements. Web pages at h t t p : / / j a v a
.sun. com/products/jdk/1.2/docs/guide/rmi/index.htnil, 1997.
47. Leendert van Doom, Martin Abadi, Mike Burrows, and Edward Webber. Secure
network objects. In Proceedings 1996 IEEE Symposium on Security and Privacy,
pages 211-221, May 1996.
48. Dennis Volpano, Cynthia Irvine, and Geoffrey Smith. A sound type system for
secure flow analysis. Journal of Computer Security, 4:167-187, 1996.
49. Edward Wobber, Martin Abadi, Michael Burrows, and Butler Lampson. Authen-
tication in the Taos operating system. ACM Transactions on Computer Systems,
12(l):3-32, February 1994.
50. Ann Wollrath, Roger Riggs, and Jim Waldo. A distributed object model for the
Java system. Computing Systems, 9(4):265-290, Fall 1996.
 Reflective Authorization Systems:
Possibilities, Benefits, and Drawbacks

Massimo Ancona, Walter Cazzola, and Eduardo B. Fernandez

Abstract. We analyze how to use the reflective approach to integrate

an authorization system into a distributed object-oriented framework.
The expected benefits from the reflective approach are: more stability
of the security layer (i.e., with a more limited number of hidden bugs),
better software and development modularity, more reusability, and the
possibility to adapt the security module with at most a few changes to
other applications. Our analysis is supported by simple and illustrative
examples written in Java.
Keywords: Authorization, Distributed Objects, Object Orientation, Reflec-
tion, Security.

1 Introduction
Security implies not only protection from external intrusions but also control-
ling the actions of internally-executing entities and the operations of the whole
software system. In this case, the interleaving between operations and data pro-
tection may become very complicated and often intractable. For this reason, se-
curity must be specified and designed in a system from its early design steps [11].
Prom another point of view
— it is very important that the security mechanisms of the application be cor-
rect and stable;
— the security code should not be mixed with the application code, otherwise
it should be very hard to reuse well-proven implementations of the security
model.
If this is not done, when a new secure application is developed the designerjim-
plementer wastes time to re-implement and to test the security modules of the
application. Moreover, security is related to: "who is allowed to do what, where
and when"; so security is not functionally part of the solution of the application
problem, but an added feature defining constraints on object interactions. From
this last remark we can think of security as a feature operating at a different
computational level and we can separate its implementation from the applica-
tion implementation.
In our opinion it is possible to exploit some typical reflection features, like sep-
aration of concerns and transparency, to split a secure system into two levels:
at the first level there are (distributed) objects cooperating to solve the sys-
tem application; at the second one, rights and authorizations for such entities
are identified, specified and mapped onto reflective entities which transparently
 36

monitor the objects of the first level and authorize the allowed access to other
objects, services, or information.
Working in this way it is possible to develop stable and reliable entities for han-
dling security. It is also possible to reuse them during system development, thus
reducing development time and costs, and increasing application level assurance.
In most systems, authorization is defined with respect to persistent data and en-
forced by the DBMS and|or operating system. Object-oriented systems define
everything as an object, some persistent some temporary, where this separation
is not visible at the application level. In these systems authorization must be
defined at the application level to take advantage of the semantic restrictions
of the information [10]. An early system (not object-oriented) (see [11], page
195), attempted this kind of control by defining programs that had predefined
and preauthorized accesses. Reflection appears as a good possibility for this type
of control because it does not separate persistent from temporary entities. The
Birlix operating system [18] used reflection to adapt its nonfunctional proper-
ties (including security) to different execution and application environments (for
more details on how satisfy nonfunctional requirements using reflection, see [21]).
In this paper we examine how to use a reflective architecture, such as those de-
scribed above, to manage the authorization aspects of an application and the
advantages and drawbacks of using such an approach.

2 Background on Reflection and Security

2.1 Reflection

Computational reflection or just reflection is defined as the activity performed

by an agent when doing computations about itself [15]. Behavioral and struc-
tural refiection are special cases which involve, respectively, agent computation
and structure (for more details see [8]).
A reflective system is logically structured in two or more levels, constituting
a reflective tower. Entities working in the base level, called base-entities or re-
flective entities, define the system basic behavior. Entities in the other levels
(meta-levels), called meta-entities, perform the reflective actions and define fur-
ther characteristics beyond the application dependent system behavior.
Each level is causally connected to adjacent levels, i.e., entities belonging to a
level maintain data structures representing (or, in reflection parlance, reifying)
the states and the structures of the entities in the level below. Any change in
the state or structure of an entity is reflected in the data structures reifying it,
and any modification to such data structures afl'ects the entity's state, structure
and behavior.
Computational reflection allows properties and functionalities to be added to
the application system in a manner that is transparent to the system itself (sep-
aration of concerns) [20]. To this respect, it is useful to consider also reflection
granularity [6], that is, the minimal entity in a software system for which a re-
flective model defines a different meta-behavior. A finer granularity allows more
 37

flexibility and modularity in the software system at tfie cost of meta-entity pro-
liferation. The reflective models considered here are: meta-ohject, and channel
reification.

Meta-Object Model. In the meta-object model, meta-entities (called meta-

objects) are objects, instances of a proper class. Each base-entity, called also
referent, can be bound to a meta-object. Such a meta-object supervises the
work of the linked referent. The meta-object model is well known in literature,
more on it can be found in [15].

Channel Reification Model. In the channel reification model [1, 2], one or
more objects, called channels are established between two interacting objects.
Each channel is characterized by a triple composed by the objects it connects
and by the kind of the meta-computation it performs.

channel = (client, server, channel-kind)

A channel kind identifies the meta-behavior provided by the channel. The kind
is used to distinguish the reflective activity to be performed: several channels
(distinguishable by the kind) can be established between the same pair of objects
at the same moment.
Each channel persists after each meta-computation, and is reused when a com-
munication characterized by the same triple is generated. The features of the
model are: method-level granularity, information continuity, channel lazy cre-
ation, and global view^. Each service request of a specific kind is trapped (shift-
up action) by the channel of the corresponding kind connecting client and server
objects, if it exists. Otherwise, such a channel is created; in either case, it then
performs its meta-computation and transmits the service request to the supplier.
The server's answer is collected and returned to the requesting object (shift-down
action).

2.2 Security: Authorization Systems

An authorization system plays a monitoring role, judging if the requests sent by
an object to another object are permissible requests. The judgment is based on
security information related to objects and subjects; where a subject represents
an entity performing or requesting an activity (i.e., an active object playing the
client role), while an object is a passive entity supplying a service (i.e., a passive
object or an active object playing the role of server). One way to model such
authorizations is the access matrix model (see [14]). In this model the autho-
rization rules are described by a bidimensional matrix indexed on subject and
objects and access to an object o is allowed to a subject s when the item (s,o) of
^ The global view concept is defined in [6] as the situation in which the meta-
computation can involve all base-entities and aspects involved in the computation
on which it reifies.
 38

the matrix contains the permitted access type. Such a model can be reaUzed by
capability lists, access control lists, or combinations of these. Formally, an autho-
rization right R for a subject s to access an object o, through message (method)
m can be written R(s)=(m,o). Here m usually represents a high-level access type,
e.g., hire an employee.
A Role-Based Access Control (RBAC) model is particularly suitable for object-
oriented systems [17], because accesses can be defined as operations defined in
the objects according to the need-to-know of the roles [9]. Because reflection
supports a fine granularity of access, its combination with RBAC can be quite
effective.

3 How We Model Security W i t h Reflection

A computational system answers questions and supports actions in some ap-
plicative domain. Following P. Maes [15] we can say that:
"A system is causally connected with its domain if the internal structures
and the domain they represent axe linked in such a way that if one of
them changes, this leads to a corresponding effect upon the other."
For example a control software is causally connected with the controlled pro-
cess if changes of the state of the physical process are reflected by changes of
the control software state and vice versa. For this purpose a control software
system incorporates structures representing the state of the controlled process.
If we consider authorization rights added to a critical control system we note
an indirect causal connection of authorization with the application domain. Au-
thorization (and more generally security) defines capabilities related to human
and/or mechanical agents concerned with the responsibility of performing tasks
that are considered critical for the controlled system behavior. Thus, an au-
thorization validation mechanism is a nonfunctional feature causally connected
with software agents and objects performing control tasks. This is our motivation
for adopting a reflective architecture for implementing authorization schemes. In
other words, security specifications correspond to nonfunctional specifications [9]
and should be implemented into a meta-level.
Using computational reflection we can separate the authorization control mech-
anism from the application. Moreover, the meta-level should be protected from
malicious intrusions and attacks from the base-level or from other processes. A
solution could consist in executing the meta-level in a different address space
as an independent process, communicating with the base-level via a mechanism
similar to the local procedure calls (LPC) of Windows NT [7]. LPC is a locally
optimized form of the well known mechanism of remote procedure call (RPC) of
Unix and other systems: LPC is a message-passing mechanism through which
clients make requests to servers and used for the server's reply. The main nec-
essary feature of the LPC mechanism is the protection domain installed around
an LPC call.
 39

3.1 Illustrative Scenario

We illustrate our ideas by using the following scenario: the system is com-
posed of several objects interacting in a chent-server manner. For security rea-
sons the services supplied by a server and its data are protected and pro-
hibited to some subjects. To support our presentation we use some stubs of
Java code. The code is tailored to the case of three clients and one server
supplying two methods, one reading the server state and the other modify-
ing it. Each client has different rights on the two supphed services. Java [3]
is not a reflective language, for our purposes we realize the reflection in a
naive way by emulating it with inheritance. Of course in this way transparency
and efficiency are compromised. However, we chose this approach for simplic-
ity and because it permits to point out the advantages and the weak points of
the reflective approach. In particular, it shows the necessity to implement the
context switch in a secure way. The complete code can be downloaded from
http://www.disi.unige.it/person/CazzolaW/OORSecurity.html.

In the next section, in order to avoid confusion with objects in object-oriented

systems, we use the term base-objects to refer either to subjects and objects of
the authorization terminology. Moreover, we call service request (or method call)
every access performed by a subject on an object. There are several reflective
architectures suitable for applying authorization rules, we consider only two of
them: the meta-object model and the channel reification model. Each of them

mo encapsulates
o 's authorizations
meta-object mo

meta-level
base-level

subject s object o

Fig. 1. Meta-object model for security

provides features suitable to handle different security aspects.

3.2 Meta-Object Approach

In the meta-object model a meta-object is associated with each object {as shown
in Fig. 1); this meta-object encapsulates the related access control list of the
 40

oa
j<l' lI'N Shell

Fig. 2. Meta-object approach: the Java structure

referred object. Each request performed by a subject s to an object o is trapped

by the related meta-object mo, which evaluates the authorization request. Then,
it either rejects the request by rising an exception, or it delivers the request to
the original destination s.
The Java architecture in shown in Fig. 2. In this case the reflective approach is
handled by the server part of the application. The client is directed to talk only
to the server. To make this approach to work, we have wrapped the server code
(the kernel of Fig. 2) into a shell which receives all client requests and performs
the shift-up action by forwarding each request to its own meta-object. Thus, a
server is composed of two parts: a kernel and a shell, related by the inheritance
hnk. As it can be deducted by inspecting the code (Listing 1), the shell waits
for a client request and forwards it to the meta-object by calling the meta-object
meta Behavior method (rows ® and ®) to check if the request is authorized or
should be rejected. Then, if the request is authorized by the meta-object, the shell
proceeds to the execution by calhng the corresponding method of the kernel (rows
® and O). The meta-object implements the authorization policy. All the work is
performed by the meta Behavior method. Method metaBehavior, when activated
for the first time, reads the referent access rights from a file and encapsulates
the referent capability list. In all other activations, method metaBehavior uses
the capability list encapsulated by the meta-object to check if the request has
to be authorized or rejected.
When a client issues the following call:

permission = server.methodRead();

a request for the execution of method Read is trapped by the shell (step O in
Fig. 2) which calls the meta-object for obtaining the authorization (step ® in
Fig. 2). The meta-object replies to the shell (step © in Fig. 2) and if authorized,
allows the kernel to return the reply to the request. Otherwise, the shell notifies
the client that a violation occurred by sending it an error message (step 0 in
Fig. 2).
 41

Listing 1 Java code for the server shell

5) p u b l i c c l a s s TheServer e x t e n d s RServer {
(2) String R E F 2 :
@ I n t e r S M O rem_obj2; / / declaration of an SMO remote object

® p u b l i c b o o l e a n m e t h o d R e a d ( i n t IDClient) {
® b o o l e a n access = r e m _ o b j 2 . m e t a B e h a v i o r { M E T H O D _ R E A D , IDClient);
® if (access) r e t u r n s u p e r . m e t h o d R e a d ( l ) ;
® e l s e r e t u r n false;
® }
® p u b l i c b o o l e a n m e t h o d W r i t e ( i n t IDClient) {
® b o o l e a n access = r e m _ o b j 2 . m e t a B e ] i a v i o r ( M E T H O D _ W R l T E , IDClient);
O if (access) r e t u r n s u p e r . m e t h o d W r i t e ( l ) ;
@ e l s e r e t u r n false;
® }
O p u b l i c s t a t i c v o i d main(Str!ng args[]) {
© System.setSecurityManager(new RMISecurityManager()); / / RMI Registration

I j Client declaration of the TheServer-i-SMO client-serving, in which we look

II up and fetch the remote object SMO.

® rem_obj2 = (InterSMO) Naming.lookup('7/SMQ"+IDServer);

& R E F = "//CS";

/ / Creation of one tnstance of the remote object TheServer

® TheServer rem_obj = n e w TheServer();

® N a m i n g . r e b i n d ( R E F , rem_obj);
® System.out.println("Waiting for some c l i e n t r e q u e s t . . . " ) ;
® }

® }

This model is suitable for implementing highly specialized role rights (specialized
per object, per subject or per service request). Moreover, each meta-object may
hold the methods necessary to modify authorization rules of the base-objects [12],
which can be called for modifying active authorizations. In this way, the devel-
oper of base-objects does not need to worry about validity of authorization
updates, which are encapsulated into the meta-object behavior. Obviously our
example suffers of efficiency penalties due to the reflective implementation in
Java.

3.3 Channel Reiflcation Approach

As we can see by inspecting the meta-object code of the previous section, the
meta-object reifies many pieces of information, and when the authorization
mechanisms become more complex, the metaBehavior code grows in complex-
ity and size, thus making the code error prone. In our example relative to the
meta-object, we use two simple access modes: allowed to read, and allowed to
write, both managed by a single meta-object. The example is quite simple but
sufficiently complex to show the increase in complexity of the code necessary to
handle all c<xses.
 42

Listing 2 Java code for the server kernel

public class RServer {
public boolean methodReacl(int i) {
System.out.println("sending methodRead");
return true;
}
public boolean methodWrite(int i) {
System.out.println("sending methodWrite")
return true;
}

channels reify and handle only

specific access modes of s too

ChannelWrite

subject s object o

Fig. 3. Channel to model access mode authorization

Another approach consists of using a reflective model with a finer granularity

than that of the meta-object model, by entrusting each access mode into a dif-
ferent reflective entity. By using the channel reification model [2] we can define a
channel kind for each existing access mode. Each element of the access matrix is
encapsulated by one or more channels (one for each value assigned to the matrix
item). For example, when a request is sent by s to o, with access mode write,
the request is trapped by the channel of kind write (ChannelWrite), established
between s and o, which vahdates it (see Fig. 3). In this way the authorization
validation process is simplified: each channel checks few authorizations and there
exists reflective entities only for base-entities requiring an authorization valida-
tion protocol. Obviously, the main drawback of this approach with respect to the
meta-object approach is a higher number of reflective entities created, however
it is possible to limit this number by merging together meta-entities showing a
similar behavior.
Due to the global view property (see [6]) of the channel reification model, each
reflective action starts from the client and involves the client, the server, and a
channel in its execution. The reflective behavior is achieved by overriding the
normal behavior of the binding|look-up between the client and the server, and
 43

Listing 3 Java code for the security meta-object

public class SMO {
static String REF2;
int idserver; / / referents id

boolean[] caplist_read = null;

booiean[] cap!ist_write = null;
public boolean metaBehavior(char MethodType, int ClientName) {
if (caplist_read = = null) {
FilelnputStream rights = new FileInputStream{"rights");
int clients, servers;
servers = rights.read{);
caplist_read = new boolean [clients = rights-read()];
caplist_write = new^ boolean [clients];
rights.skip(idserver*{2*clients+l));
for (int i=0; i<clients;H-+) {
cap]ist_read[i] = {(rights.read() = = "R")?true.false);
capiist_write[i] = ((rights.read() = = "W")?trne.false);
}
return ((method_type = = "R")?caplist-reacl[ClieiitName]:caplist_write[ClientName]);
}
public static void main(String argsf]) {
idserver = args[0];
System.setSecurityManager(new RMISecnrityManager());
REF2 = "//SMO";
SMO rem_obj2 = new SMO();
Naming.rebind(REF2, rem_obj2);
System.out.println("Security meta-object waits for some request. . ");
}

the remote method invocation in order to divert the request from the server to
the channel. The chent is composed of three layers (see Fig. 4). The inner layer
supplies the operations necessary to communicate with the server. The central
layer encapsulates all operations of the inner layer and uses them to communi-
cate with the right channel. The outer layer defines the behavior of the client.
The server code corresponds exactly to the kernel code of the previous approach.
There is no need to wrap it because the server receives requests that have been
previously filtered by the channels, it cannot be accessed directly by the clients.
The above splitting of responsibility of the validation mechanism, makes the
code of a channel simpler and more modular than that of the meta-object. Op-
erations channelRead and channelWrite are very similar, they differ only in the
encapsulated information and the use they make of it. In our example channels
and meta-objects are similar because the validation phase is simple; however, it
should be noted that the amount of information encapsulated and managed by
a single channel has decreased.
When a client issues the call (see Fig. 4):

permission = server.methodRead();
 44

Listing 4 Java code for the inner layer

public class operations {
int IDClieiit;

public static Object rem_server:

/ / Here the remote channel object is bound to the URL

public Object Binding(String str) {

return Naming.lookup("//" + str);
}
/ / This function imtokes the remote methods in TheServer

public Object rmi(int method, int ClientName) {

if (method == 0) return ((InterCS)rem_server).Method-Read(ChentName)i
else return ((InterCS)rem_server).Method-Write(ChentName);
}

Central Layer
^Outer Layer

Fig. 4. Java structure for the channel reification approach

the control is dispatched to the central layer of the client (step O in Fig. 4) which
determines the right channel to be invoked on the basis of the request kind. The
request is then forwarded to the channel while the client idles until a reply is
sent back to it (step @ in Fig. 4). The channel validates the request and, if legal,
forwards it to the server (step @ in Fig. 4). When the channel receives the reply
from the server, it forwards it back to the client (step O in Fig. 4). The server
is unaware of the validation process: it executes only filtered requests. Thus a
good separation of concerns and implementation modularity is achieved.

Security models based on communication flow [5] can be easily modeled by the
channel reification model: in particular, models based on the concept of flow
such as those of Escort of Scout [19] and Corps [16]. A path is a first class object
encapsulating data flowing through a set of modules. A path is a logical channel
 45

Listing 5 Java code for the central layer

public class reflectiveLayer extends operations {

Boolean Right;
private InterCCRead rem-chr;
private InterCCWrite rem-chw;
/ / Here the remote channel object is prepared to bind the URL
j j (see the Operations class for the actual binding )
public Object Binding(String str) {
rem-chr = (InterCCRead) super.Binding("CCR"+IDClient + "-"+lDServer);
rem_chw = (InterCCWrite) super.Binding("CCW"+IDCIient+"-"+IDServer);
return null;
}
/ / Here the rmi method invokes the metaBehavior function in the Channel
j j classes, where the checking about client permissions is performed

public Object rmi{int method) {

if (method—=0) Right — rem_chr.metaBehavior(method);
if (method=:=l) Right = rem_chw.metaBehavior(method);
return Right;
}
}

made up of data flow connecting several modules. Other two security mecha-
nisms adopted in Escort are filters and protection domains. A filter restricts the
interface between two adjacent modules. However, filters include no mechanism
to ensure that a module does not bypass the interface by directly accessing the
memory of the other module. A protection domain is a boundary drawn between
a pair of modules to ensure that the mutual access is performed only through the
defined interface. Filters and protection domains may be modeled respectively
by standard reflective channels and protected channels, i.e., channels executing
in a different address space. The concept of path is more complex and requires
an extension of the channel reification mechanism. A channel controlling a path
may be obtained by piping or composing the channels controlling the sequence
of modules forming a path or by defining a complex channel successively control-
ling the interface of a sequence of modules in a path. Another approach could be
based on a three-level reflective tower, but this approach increases the system
complexity without significant advantages.

4 Evaluation

Using computational reflection to include an authorization mechanism into a

software system offers many advantages during software development. We can
specify, develop, implement and test the modules which implement authorization
mechanisms separately from the rest of the application. In this way we encourage
reuse and improve software stability. Moreover, the authorization process is hid-
 46

Listing 6 Java channel code

public class chaiinelRead {

int IDClient, IDServer; Boolean result = null; String REF; InterCS rem-objS;
/ / Method Constructor with two parameters rised to identify channel's client and server
public channelRead(int c, int s) {
IDClient = c; IDServer ~ s;
}
public Boolean metaBehavior(int MethoclType) {
if (result =:— null) { / / we have to read access matrix in order to initialize channel

}"
if (cell —— 34) { {I 34 is the int code for "R"
System.out.println("\nYou have the READ permission !");
return rem_objS.Method_Read();
else {
System.out.println("\nYou don't have the READ permission ! !");
return Boolean.FALSE;
}
}
public static void main(String args[]) {
cliannelRead rem_obj;
/ / creation and installation of the security manager
System.setSecurityManager( new RMISecurityManager{));

/ / Here the remote TheServer object is bound to the URL

rem_objS = ( InterCS ) Naming.lookup("//CS");
/ / this is the server set-up for TheClient-chajinelRead system
REF = 'V/CCR"+args[0]4-args[l];
/ / creation of an instance of the remote class
rem_obj = new channelRead(Integer.parseInt(args[0]), Integer.parselnt(args[l]));

/ / creation of the registry-remote object binding

Naming.rebind(REF, rem_obj};
System.out-println("Reaiiing channel["+args[0] + " ,*'+args[l] + "] waits for some requests");
}

den to the application entities, thus the code of such entities is simpUfied. While
this can be accomplished by current DBMS authorization systems, we are now
controlling access to all executing entities, not just the application's persistent
data.
The advantages from the security point of view are that only the entities per-
forming the validation of the authorizations of an object know its authorizations
constraints. In this way authorization leaking is minimized.
A first drawback is that flexibility has a cost in terms of efficiency due to repeated
method activation. The Achille's heel of using reflection to realize security could
be its implementation mechanism. In fact, as it is stressed by our examples, each
reflective approach has a mechanism, which permits to forward each request to
meta-entities. These trap actions (shift-up and shift-down) are critical actions. It
is possible for a malicious user to intercept the trap and to hijack the request to
 47

another entity bypassing the authorization system and illegally authorizing each
request. For this reason it is important to protect the meta-level and the access
to it and for the same reason it is important to hmit the meta-entities commu-
nication and, thus, using a reflective model having the global view feature. One
way to avoid this attack consists of implementing the whole meta-computation,
and in particular the trap actions, as protected routines. Fiom another point of
view this drawback can be converted into an advantage of the model with respect
to standard authorization models, because it minimizes the vulnerable points to
known system locations that can be more efficiently protected. The possibility
of using different address spaces (e.g., different processes) to implement the two
reflective layers with kernel system intervention for exchanging information rep-
resents an improvement to the security of the complete system.
Some recent proposals to control actions of Java applets have similar purposes
to our proposal. In those systems, e.g., in [13], execution domains are created
and enforced for specific applications using downloaded content. However, the
objective of these approaches is to control access to operating system resources,
e.g., files, memory spaces, ..., not to control high-level actions between objects.

5 Conclusions

Reflection offers several advantages when used to model authorization mecha-

nisms. Its main advantages are separation of concerns and modularity. Autho-
rization mechanisms can be designed within the application from early develop-
ment stages, but, at the same time, they can be maintained separate both from
the logical and implementative point of view. This fact improves reusability of
both functional and authorization software and supports an independent testing
of both. More important, it controls access to all executing entities, not just to
persistent data.
Another advantage is the ability of implementing a protection layer around the
authorization software, thus making the system more robust to unauthorized
attempts to change role rights. Obviously, there are also drawbacks: the first is
a reduced execution efficiency; flexibility costs in efficiency. The second problem
could be presented by the protection mechanism around the authorization layer
(meta-level). Running it in a different address space may make programs too in-
efficient for most applications. Thus, more efficient protection mechanisms, not
performing a complete context switching, should be designed. Hardware capa-
bility systems appear promising for this purpose.
Moreover, the existence of such a protection makes more understandable, to
malicious users, where to address attacks to the fortress and easier to discover
Achilles' heels.
Future developments are represented by complete workflow authorizations com-
bining specific sequences of service requests: they require that the meta-entities
controlling the activated server objects, interact with each other for discriminat-
ing legal sequences from illegal ones. More complex authorization schemes may
require the introduction of meta-meta-levels, i.e., to raise the reflective tower
and the system complexity.
 48

Finally, some prototyping of the proposed architecture and practical experiments

will improve the understanding of the role played by reflection in t h e implemen-
tation of authorization systems of high assurance. In particular, its possible use
to control the actions of downloaded content would be of high practical interest.
Another possible use is for the control of information flow in object-oriented
systems as proposed in [4].

Acknowledgments
A preliminary version of this work appears in the proceedings of the 1*' E C O O P
Workshop on Distributed Object Security, pages 35-39, Belgium, July 1998.

References
[1] Massimo Ancona, Walter Cazzola, Gabriella Dodero, and Vittoria Gi-
anuzzi. Channel Reification: a Reflective Approach to Fault-Tolerant
Software Development. In OOPSLA '95 (poster section), page 137,
Austin, Texas, USA, on 15th-19th October 1995. ACM. Available at
http://www.disi.unige.it/person/CazzolaW/references.html.
[2] Massimo Ancona, Walter Cazzola, Gabriella Dodero, and Vittoria Gianuzzi.
Channel Reification: A Reflective Model for Distributed Computation. In proceed-
ings of IEEE International Performance Computing, and Communication Confer-
ence (IPCCC'98), 98CH36191, pages 32-36, Phoenix, Arizona, USA, on 16th-18th
February 1998. IEEE.
[3] Ken Arnold and James Gosling. The Java Programming Language. The Java Se-
ries ... from the Source. Addison-Wesley, Reading, Massachussetts, second edition,
December 1997.
[4] Elisa Bertino, Sabrina De Capitani di Vimercati, Elena Ferrari, and Pierangela
Samarati. Exception-Based Information Flow Control in Object-Oriented Sys-
tems. ACM TYansactions on Information and System Security (TISSEC), 1(1),
November 1998.
[5] W. E. Boebert and R. Y. Kain. A Pratical Alternative to Hierarchical Integrity
Policies. In proceedings of 8th National Computing Security Conference, Gaithers-
burg, October 1985.
[6] Walter Cazzola. Evaluation of Object-Oriented Reflective Models. In proceed-
ings of ECOOP Workshop on Reflective Object-Oriented Programming and Sys-
tems (EWROOPS'98), in 12th European Conference on Object-Oriented Pro-
gramming (ECOOP'98), Brussels, Belgium, on 20th-24th July 1998. Available at
http://www.disi.unige.it/person/CazzolaW/references.html.
[7] Helen Custer. Inside Windows NT. Microsoft Press, Redmond, WA, 1993.
[8] Prangois-Nicola Demers and Jacques Malenfant. Reflection in Logic, Functional
and Object-Oriented Programming: a Short Comparative Study. In proceed-
ings of the workshop section, in IJCAr95 (International Join Conference on AI),
Montreal, Canada, August 1995.
[9] Eduardo B. Fernandez and J. C. Hawkins. Determining Role Rights from Use
Cases. In proceedings of the 2nd ACM Workshop on Role Based Access Control
(RBAC'97), pages 121-125, November 1997.
 49

[10] Eduardo B. Fernandez, Maria M. Larrondo-Petrie, and Ehud Gudes. A Method-

Based Authorization Model for Object-Oriented Databases. In proceedings of
the OOPSLA'93 Workshop on Security in Object-Oriented Systems, pages 70-79.
ACM, 1993.
[11] Eduardo B. Fernandez, Rita C. Summers, and Christopher Wood. Database Se-
curity and Integrity. Addison-Wesley, Reading, Massachusetts, 1981.
[12] Michael A. Harrison, Walter L. Ruzzo, and Jeffrey D. Ullman. Protection in
Operating Systems. Communication of the ACM, 19(8);461-471, August 1976.
[13] Trent Jaeger, Nayeen Islam, Rangachari Anand, Atul Prakash, and
Jochen Liedtke. Flexible Control of Downloaded Executable Content.
http://www.ibm.com/Java/education/flexcontrol, 1997.
[14] Butler W. Lampson. Protection. Operating System Review, 8(l):18-34, January
1974. Reprint.
[15] Pattie Maes. Concepts and Experiments in Computational Reflection. In proceed-
ings of OOPSLA '87, volume 22 of Sigplan Notices, pages 147-156. ACM, October
1987.
[16] Edwin Menze, F. Reynolds, and F. Travostino. Programming with System Re-
sources in Support of Real-Time Distributed Applications. In proceedings of the
1996 IEEE Workshop on Object-Oriented Real-Time Dependable Systems, pages
36-45, Laguna Beach, Ca, February 1996. IEEE.
[17] Ravi Sandhu, Edward J. Coyne, Hal L. Feinstein, and Charles E Youman. Role-
Based Access Control Models. Computer, 29(2):38-47, Febriiar\' 1996.
[18] Susann Sonntag, Hermann Hartig, Oliver Kowalski, Winfried Kiihnhauser, and
Wolfang Lux. Adaptability Using Reflection. In proceedings of the 27th Annual
Hawaii International Conference on System Sciences, pages 383-392, 1994.
[19] Oliver Spatscheck and Larry L. Peterson. Escort: A Path-Based OS Security
Architecture. Technical Report TR-97-17, Department of Computer Science, The
University of Arizona, Tucson, AZ 85721, November 1997.
[20] Robert J. Stroud. Transparency and Reflection in Distributed Systems. ACM
Operating System Review, 22:99-103, April 1992.
[21] Robert J. Stroud and Zhixue Wu. Using Metaobject Protocols to Satisfy Non-
Functional Requirements. In Chris Zimmerman, editor. Advances in Object-
Oriented Metalevel Architectures and Reflection, chapter 3, pages 31-52. CRC
Press, Inc., 2000 Corporate Blvd.,N.W., Boca Raton, Florida 33431, 1996.
 Abstractions for Mobile Computation
LucaCardelli

Abstract. We discuss the difficulties caused by mobile computing and mobile

computation over wide-area networks. We propose a unified framework for over-
coming such difficulties.

1 Introduction
The Internet and the World-Wide-Web provide a computational infrastructure that
spans the planet. It is appealing to imagine writing programs that exploit this global in-
frastructure. Unfortunately, the Web violates many familiar assumptions about the be-
havior of distributed systems, and demands novel and specialized programming
techniques. In particular, three phenomena that remain largely hidden in local-area net-
work architectures become readily observable on the Web:
• (A) Virtual locations. Barriers are erected between mutually distrustful admin-
istrative domains. Therefore, a program must be aware of where it is, and of how
to move or communicate between different domains. The existence of separate
administrative domains induces a notion of virtual locations and of virtual dis-
tance between locations.
• (B) Physical locations. On a planet-size structure, the speed of light becomes
tangible. For example, a procedure call to the antipodes requires at least 1/10 of
a second, independently of future improvements in networking technology. This
absolute lower bound to latency induces a notion of physical locations and phys-
ical distance between locations.
• (C) Bandwidth fluctuations. A global network is susceptible to unpredictable
congestion and partitioning, which results in fluctuations or temporary interrup-
tions of bandwidth. Moreover, mobile devices may perceive bandwidth changes
as a consequence of physical movement. Programs need to be able to observe
and react to these fluctuations.
These features may interact among themselves. For example, bandwidth fluctuations
may be related to physical location because of different patterns of day and night net-
work utilization, and to virtual location because of authentication and encryption across
domain boundaries. Virtual and physical locations are often related, but need not coin-
cide.
In addition, another phenomenon becomes unobservable on the Web:
• (D) Failures. On the Web, there is no practical upper bound to communication
delays. In particular, failures become indistinguishable from long delays, and
thus undetectable. Failure recovery becomes indistinguishable from intermittent
 52

connectivity. Furthermore, delays (and, implicitly, failures) are frequent and un-
predictable.
These four phenomena determine the set of ohservables of the Web: the events or
states that can be in principle detected. Observables, in turn, influence the basic building
blocks of computation. In moving from local-area networks to wide-area networks, the
set of observables changes, and so does the computational model, the programming
constructs, and the kind of programs one can write. The question of how to "program
the Web" reduces to the question of how to program with the new set of observables
provided by the Web.
At least one general technique has emerged to cope with the observables character-
istic of a wide-area network such as the Web. Mobile computation is the notion that run-
ning programs need not be forever tied to a single network node. Mobile computation
can deal in original ways with the following phenomena:
• (A) Virtual locations. Given adequate trust mechanisms, mobile computations
can cross barriers and move between virtual locations. Barriers are designed to
impede access, but when code is allowed to cross them, it can access local re-
sources without the impediment of the barrier.
• (B) Physical locations. Mobile computations can move between physical loca-
tions, turning remote calls into local calls, and thus avoiding the latency limit.
• (C) Bandwidth fluctuations. Mobile computations can react to bandwidth fluc-
tuations, either by moving to a better-placed location, or by transferring code
that establishes a customized protocol over a connection.
• (D) Failures. Mobile computations can run away from anticipated failures, and
can move around presumed failures.
Mobile computation is also strongly related to recent hardware advances, since
computations move implicitly when carried on portable devices. In this sense, we can-
not avoid the issues raised by mobile computation: more than an avant-garde software
technique, it is an existing hardware reality.
In this paper, we discuss mobile computation at an entirely informal level; formal
accounts of our framework can be found in [ 13]. In Section 2 we describe the basic char-
acteristics of our existing computational infrastructure, and the difficulties that must be
overcome to use it effectively. In Section 3 we review existing ways of modeling dis-
tribution and mobility. In Section 4 we introduce an abstract model, the ambient calcu-
lus, that attempts to capture fundamental features of distribution and mobility in a
simple framework. In Section 5, we discuss applications of this model to programming
issues, including a detailed example and a programming challenge.

2 Three Mental Images

We begin by comparing and contrasting three mental images; that is, three abstracted
views of distributed computation. From the differences between these mental images
we derive the need for new approaches to global computation.
 53

2.1 Local Area Networks

The first mental image corresponds to the now standard, and quickly becoming obso-
lete, model of computation over local area networks.
When workstations and PCs started replacing mainframes, local networks were in-
vented to connect autonomous computers for the purpose of resource sharing. A typical
local area network consists of a collection of computers of about the same power (with-
in a couple of hardware generations) and of network links of about the same bandwidth
and latency. This environment is not always completely uniform: specialized machines
may operate as servers or as engineering workstations, and specialized subnetworks
may offer optimized services. Still, by and large, the structure of a LAN can be depicted
as the uniform network of nodes (computers) and links (connections) in Mental
Image 1;

Administrative Domain

Mental Image 1: Local Area Network

A main property of such a network is its predictability. Communication delays are
bounded, and processor response times can be estimated. Therefore, link and process
failures can be detected by time-outs and by "pinging" nodes.
Another important property of local area networks is that they are usually well-ad-
ministered and, in recent times, protected against attack. Network administrators have
the task of keeping the network running and protect it against infiltration. In the picture,
the boundary line represents an administration domain, and the flames represent the
protection provided by a firewall. Protection is necessary because local area networks
are rarely completely disconnected: they usually have slower links to the outside world,
which are however enough to make administrators nervous about infiltration.
The architecture of local area networks is very different from the older, highly cen-
tralized, mainframe architecture. This difference, and the difficulties implied by it, re-
sulted in the emergence of novel distributed computing techniques, such as remote-
procedure-call, client-server architecture, and distributed object-oriented programming.
The combined aim and effect of these techniques is to make the programming and ap-
plication environment stable and uniform (as in mainframes). In particular, the network
topology is carefully hidden so that any two computers can be considered as lying one
 54

logical step apart. Moreover, computers can be considered immobile; for example, they
usually preserve their network address when physically moved.
Even in this relatively static environment, the notion of mobility has gradually ac-
quired prominence, in a variety of forms. For example;
• Control mobility. During an RPC (Remote Procedure Call) or RMI (Remote
Method Invocation) call, a thread of control is thought of as moving from one
machine to another and back.
• Data mobility. In RPC/RMI, data is linearized, transported, and reconstructed
across machines.
• Link mobility. The end-points of network channels, or remote object proxies,
can be transmitted.
• Object mobility. For load balancing purposes, objects can be moved between
different servers.
• Remote Execution. Computations can be shipped for execution on a server.
(This is an early version of code mobility, proposed as an extension of RPC.
[35])
In recent years, distributed computing has been endowed with greater mobility
properties and easier network programming. Techniques such as Object Request Bro-
kers have emerged to abstract over the location of objects providing certain services.
Code mobility has emerged in Tel and other scripting languages to control network ap-
plications. Agent mobility has been pioneered in Telescript [37], aimed towards a uni-
form (although wide-area) network of services. Closure mobility (the mobility of active
and connected entities) has been investigated in Obliq [11].
In due time, local-area-network techniques would have smoothly and gradually
evolved towards deployment on wide-area networks, e.g. as was explicitly attempted by
the CORBA effort. But, suddenly, a particular wide-aiea network came along that rad-
ically changed the fundamental assumptions of distributed computing and its pace of
progress; the Web.

2.2 Wide Area Networks

Global computing evolved over the span of a few decades in the form of the Internet.
But it was not until the emergence of the Web that the peculiar characteristics of the In-
ternet were exposed in a way that anybody could verify with just a few mouse clicks.
For clarity and simplicity we will refer to the Web as the primary global information
infrastructure, although it was certainly not the first one.
We should remember that the notions of a global address space and of a global file
system have been popular at times as an extension of mainframe architecture to wide-
area networks. The first obvious feature of the Web is that, although it forms a global
computational resource, it is nothing like a global mainframe. The Web has no single
reliable component, but it also has no single failure point; it is definitely not the central-
ized all-powerful mainframe of 1950's science fiction novels that could be shut off by
attacking its single "brain"'. Fortunately, the fact that the Web is not a mainframe is not
a big concern; we have already successfully tackled distributed computing based on
LANs.
 55

More distressing is the fact that the Web does not behave like a LAN either. Many
proposals have emerged along the lines of extending LAN concepts to a global environ-
ment; that is, in turning the Internet into a distributed address space, or a distributed fde
system. However, since the global environment does not have the stability properties of
a LAN, this can be achieved only by introducing redundancy (for reliability), replica-
tion (for quality of service), and scalability (for management) at many different levels.
Things might have evolved in this direction, but this is not the way the Web came to be.
The Web is, almost by definition, unreliable, unpredictable, and unmanageable as a
whole, and was not designed with LAN-like guarantees of service.

Mental Image 2: Wide Area Network (for example, the Web)

Therefore, the main problem with the Web is that it is not just a big LAN, other-
wise, modulo issues of scale, we would already know how to deal with it. There are sev-
eral ways in which the Web is not a big LAN, and we will describe them shortly. But
the fundamental reason is that, unlike a LAN, the Web is not centrally administered. In-
stead, it is a dynamic collection of countless independent administrative domains, all
widely different and mutually distrustful. This is represented in Mental Image 2.
In that picture, computers differ greatly in power and availability, while network
links differ greatly in capacity and reliability. Large physical distances have visible ef-
fects, and so do time zones. The architecture of a wide-area network is yet again funda-
mentally different from that of a local area network. Most prominently, the network
topology is dynamic and non-trivial. Computers become intrinsically mobile: they re-
quire different addressing when physically moved across administrative boundaries.
Techniques based on mobility become more important and sometimes essential. For ex-

' Still, a single faulty routing configuration file spread over the Internet in July 1997,
causing the disappearance of a large number of Internet domains. In this case, the
vulnerable "brain" was the collection of Internet routers.
 56

ample, mobile Java applets provided the first disciplined mechanism for running code
able to (and allowed to) systematically penetrate other people's firewalls. Countless
projects have emerged in the last few years with the aim of supporting mobile compu-
tation over wide areas, and are beginning to be consolidated.
At this point, our architectural goal might be to devise techniques for managing
computation over an unreliable collection of fcir-flung computers. However, this is not
yet the full picture. Not only are network links and nodes widely dispersed and unreli-
able; they are not even liable to stay put, as we discuss next.

2.3 Mobile Computing

A different global computing paradigm has been evolving independently of the Web.
Instead of connecting together all the LANs in the world, another way of extending the
reach of a LAN is to move individual computers and other gadgets from one LAN to
another, dynamically.
We discussed in the Introduction how the main characteristics of the Web point to-
wards mobile computation. However, that is meant as mobile computation over a fixed
(although possibly flaky) network. A more interesting picture emerges when the very
components of the network can move about. This is the field of mobile computmg. To-
day, laptops and personal organizers routinely move about; in the future entire networks
will go mobile (as in IBM's Personal Area Network). Existing examples of this kind of
mobility include: a smart card entering a network computer slot; an active badge enter-
ing a room; a wireless PDA or laptop entering a building; a mobile phone entering a
phone cell.
We could draw a picture similar to Mental Image 1, but with mobile devices mov-
ing within the confines of a single LAN. This notion of a dynamic LAN is a fairly minor
extension of the basic LAN concepts, and presents few conceptual problems (wireless
LANs are already common). A much more interesting picture emerges when we think
of mobile gadgets over a WAN, because administrative boundaries and multiple access
pathways then interact in complex ways, as anybody who travels with a laptop knows
all too well.

m
Tjs\ ® /^

Mental Image 3: Mobile Computing

Mental Image 3 focuses on two domains: the United States and the European
Union, each enclosed by a political boundary that regulates the movement of people and
 57

computers. Within a political boundary, private companies and public agencies may
further regulate the flow of people and devices across their doors. Over the Atlantic we
see a third domain, representing Air France flight 81 travelling from San Francisco to
Paris. AF81 is a very active mobile computational environment: it is full of people
working with their laptops and possibly connecting to the Internet through airphones.
(Not to mention the hundreds of computers that control the airplane and let it commu-
nicate with ground stations.)
Abstracting a bit from people and computation devices, we see here a hierarchy of
boundaries that enforce controls and require permissions for crossing. Passports are re-
quired to cross political boundaries, reservations are required for restaurants, tickets are
required for airplanes, and special clearances are required to enter (and exit!) agencies
such as the NSA. Sometimes, whole mobile boundaries cross in and out of other bound-
aries and similarly need permissions, as the mobile environment of AF81 needs permis-
sion to enter an airspace. On the other hand, once an entity has been allowed across a
boundary, it is fairly free to roam within the confines of the boundary, until another
boundary needs to be crossed.

2.4 General Mobility

We have described two different notions of mobility. The first, mobile computation, has
to do with virtual mobility (mobile software). The second, mobile computing, has to do
with physical mobility (mobile hardware). These two fields are today almost discon-
nected, the first dominated by a software community, and the second dominated by a
hardware community. However, the borders between virtual and physical mobility are
fuzzy, and eventually we will have to treat all kinds of mobility in a uniform way. Here
are two examples where the different forms of mobility interact.
The first example is one of virtual mobility achieved by physical means. Consider
a software agent in a laptop. The agent can move by propagating over the network, but
can also move by being physically transported with the laptop from one location to an-
other. In the first case, the agent may have to undergo security checks when it crosses
administration domains. In the second case the agent may have to undergo security
checks (e.g. virus detection) when the laptop is physically allowed inside a new admin-
istrative domain. Do we need two completely separate security infrastructures for these
two cases, or can we somehow find a common principle? A plausible security policy for
a given domain would be that a physical barrier (a building door) should provide the
same security guarantees as a virtual barrier (a firewall).
The second example is one of physical mobility achieved by virtual means. Soft-
ware exists that allows remote control of a computer, by bringing the screen of a remote
computer on a local screen. The providers of such software may claim that this is just
as good as moving the computer physically, e.g. to access its local data. Moreover, if
the remote computer has a network connection, this is also equivalent to "stringing
wire" from the remote location, since the remote network is now locally accessible. For
example, using remote control over a phone line to connect from home to work where
a high-bandwidth Internet connection is available, is almost as good as having a high-
bandwidth Internet connection brought into the home.
The other side of the coin of being mobile is of becoming disconnected or intermit-
 58

tently connected. Even barring flaky networks, intermittent connectivity can be caused
by physical movement, for example when a wireless user moves into some form of
Faraday cage. More interestingly, intermittent connectivity may be caused by virtual
movement, for example when an agent moves in and out of an administrative domain
that does not allow communication. Neither case is really a failure of the infrastructure;
in both cases, lack of connectivity may in fact be a desirable security feature. Therefore,
we have to assume that intermittent connectivity, caused equivalently by physical or
virtual means, is an essential feature of mobility.
In the future we should be prepared to see increased interactions between virtual
and physical mobility, and we should develop frameworks where we can discuss and
manipulate these interactions.

2.5 Barriers and Action-at-a-Distance

The unifying difficulty in both mobile computing and mobile computation is the prolif-
eration of barriers, and the problems involved in crossing them. This central difficulty
implies that we must regard barriers as fundamental features of our computational mod-
els. This seems contrary to the usual trend.
Access barriers have arisen many times in the history of computing, and one of the
main tasks of computer science has been to "abstract them away", often by the prover-
bial extra level of indirection. For example, physical memory boundaries are circum-
vented by virtual memory; address space boundaries are circumvented by network
proxies; firewall boundaries are circumvented by secure tunnels and agent sandboxing.
Unfortunately, when barriers are not purely technological it is not possible to complete-
ly abstract them away. The crossing of administrative barriers must be performed by ex-
plicit bureaucratic operations, such as exhibiting equipment removal passes and export
licences.
Therefore, administrative barriers constitute a nasty fundamental change in the way
we compute. Let's review some historical examples.
In the early days of the Internet, any computer could talk to any other computer by
knowing its IP number. We can now forget about flat IP addressing and transparent
routing: routers and firewalls effectively hide certain IP addresses from view and make
them unreachable by direct means.
In the early days of programming languages, people envisioned a universal address
space in which all programs would live and share data, possibly with world-wide gar-
bage-collection, and possibly with strong typing to guarantee the integrity of the point-
ers. We can now forget about universal addressing: although pointers are allowed
across machines on a LAN (by network proxies), they are generally disallowed across
firewalls. Similarly, we can forget about transparent distributed object systems: some
network objects will be kept well hidden within certain domains, and reaching them will
require effort.
In the early days of mobile agents, people envisioned agents moving freely across
the network on behalf of their owners. We can now forget about free-roaming: if sites
do not trust agents they will not allow them in. If agents do not trust sites to execute
them fairly, they will not want to visit them.
In general, we can forget about the notion of action-at-a-distance computing: the
 59

idea that resources are available transparently at any time, no matter how far away. In-
stead, we have to get used to the notion that movement and communication are step-by-
step activities, and that they are visibly so: the multiple steps involved cannot be hidden,
collapsed, or rendered atomic.
The action-at-a-distance paradigm is still prevalent within LANs, and this is anoth-
er reason why LANs are different from WANs, where such an assumption cannot hold.
2.6 Why a WAN is not a big LAN
We have already discussed in the Introduction how a WAN exhibits a different set of
observable than a LAN. But could one emulate a LAN on top of a WAN, restoring a
more familiar set of observables, and therefore a more familiar set of programming
techniques? If this were possible, we could then go on and program the Internet just like
we now program a LAN.
To turn a WAN into a LAN we would have to hide the new observables that a WAN
introduces, and we would have to reveal the observables that a WAN hides. These tasks
ranges from difficult, to intolerable, to impossible. Referring to the classification in the
Introduction, we would have to achieve the following.
(A) Hiding virtual locations. We would have to devise a security infrastructure that
makes navigation across multiple administration domains painless and transparent
(when legitimate). Although a great deal of cryptographic technology is available, there
may be impossibility results lurking in some corners. For example, it is far from clear
whether one can in principle guarantee the integrity of mobile computations against
hostile or unfair servers [33]. (This can be solved on a LAN by having all computers
under physical supervision.)
(B) Hiding physical locations. One cannot "hide" the speed of light; techniques
such as caching and replication may help, but they cannot fool processes that attempt to
perform long-distance real-time control and interaction. In principle, one could make
all delays uniform, so that programs would not behave differently in different places.
Ultimately this can be achieved only by slowing down the entire infrastructure, by em-
bedding the maximal propagation delay in all communications. (This would be about 1/
10 of a second on the surface, but would grow dramatically as the Web is extended to
satellite communication, orbital stations, and further away.)
(C) Hiding bandwidth fluctuations. It is possible to introduce service guarantees
in the networking infrastructure, and therefore eliminate bandwidth fluctuations, or re-
duce them below certain thresholds. However, in overload situations this has the only
effect of turning network congestion into access failures, which brings us to the next
point.
(D) Revealing failures. We would have to make failures as observable as on a
LAN. This is where we run into fundamental trouble. A basic result in distributed sys-
tems states that we cannot achieve distributed consensus (such as agreeing on which
nodes have failed) in a system consisting of a collection of asynchronous processes
[19]. The Web is such a system: we can make no assumption about the relative speed
of processors (they may be overloaded, or temporarily disconnected), about the speed
of communication (the network may be congested or partitioned), about the order of ar-
rival of messages, or even about the number of processes involved in a computation. In
 60

these circumstances, it is impossible to detect the failure of processors or of network

nodes or links; any consensus algorithm can be delayed indefinitely. The common par-
tial solutions for this unsolvable problem are to dictate some degree of synchrony and
failure detection. These solutions work well on a LAN, but they seem unlikely to apply
to WANs simply because individual users may arbitrarily decide to turn off their pro-
cessors without warning, or take them into unreachable places. Other partial solutions
involve multiple-round broadcast-based probabilistic algorithms [9] which might be ex-
pensive on a WAN in terms of communication load, and would be subject to light-speed
delays. Moreover, it is difficult to talk about the failure of processors that are invisible
because they are hidden behind firewalls, and yet take part in computations. Therefore,
it seems unlikely that techniques developed to deal with asynchrony in operating sys-
tems and LANs can be successfully applied to a WAN such as the Web in full general-
ity. The Web is an inherently asynchronous system, and the impossibility result of [19]
applies with full force.
In summary: task (A) may be unsolvable for mobile code; in any case, a non-zero
amount of bureaucracy will always be required; task (B) is only solvable (in full) by in-
troducing unacceptable delays; task (C) can be solved in a way that reduces it to (D);
task (D) is unsolvable in principle, while probabilistic solutions run into task (B).

2.7 WAN Postulates

We summarize this section by a collection of postulates that capture the main properties
of the reality we are interested in modeling:
• 1) Separate locations exist.
• 2) Since different locations have different properties, both people and programs
will want to move between them.
• 3) Barriers to mobility will be erected to preserve certain properties of certain
locations.
• 4) Some people and some programs will still need to cross those barriers.
The point of these postulates is to stress that mobility and barrier crossing are inevitable
requirements of our current and future computing infrastructure.
The observables that are characteristic of wide-area networks have the following
implications:
• Distinct virtual locations:
- Are observed because of the existence of distinct administrative domains, which
are produced by the inevitable existence of attackers.
- Preclude the unfettered execution of actions across domains. Preclude continuous
connectivity.
- Require a security model.
• Distinct physical locations:
- Are observed, over large distances, because of the inevitable latency limit given
by the speed of light.
 61

- Preclude instantaneous action at a distance.

- Require a mobility model.
• Bandwidth fluctuations (including hidden failures):
- Are observed because of the inevitable exercise of free will by network users,
both in terms of communication and movement.
- Preclude reliance on response time.
- Require an asynchronous communication model.

3 Modeling Mobility
Section 2 was dedicated to showing that the reality of mobile computation over a WAN
does not fall into familiar categories. Therefore, we need to invent a new paradigm, or
model, that can help us in understanding and eventually in taking advantage of this re-
ality.
Since the Web is, after all, a distributed system, it is worth reviewing the existing
literature on models of distributed systems to see if there is something there that we can
already use. Readers who are not interested or experienced in models of concurrency,
may skip ahead to Section 4.

3.1 Formalisms for Concurrency, Distribution, and Security

The Ti-calculus [30], along with its variations, is a prominent model of concurrency, and
is the starting point for our work. It is based on the notion of processes communicating
over channels, with the ability to create new channels and to exchange channels over
channels.
We spend some time discussing why some of the basic assumptions of the Tt-cal-
culus and of other concurrent formalisms do not satisfy our particular needs.
Static Connectivity
Some early paradigms for concurrency, such as Actors [3], allowed highly dynamic sys-
tems. However, the main formalized descriptions of concurrency began by considering
only static connectivity. This is the case for Petri Nets [10], for Hoare's Communicating
Sequential Processes (CSP) [24], and for Milner's Calculus of Communicating Systems
(CCS) [28]. In CCS, in particular, the set of communication channels that a processes
has available does not change during execution. In some versions of CSP (and in Occam
[26]) the set of processes cannot change either.
These models are insufficient for modeling mobility in a general sense. Instead,
they best characterize the situation in Mental Image I, where the only form of mobility
is the mobility of data over communication channels.
Dynamic Connectivity
The Ji-calculus is an extension of CCS where channels can be transmitted over other
channels, so that a process can dynamically acquire new channels. Channel transmis-
sion (also called channel mobility) is a powerful extension of the basic communication
model. The transmission of a channel over another channel gives the recipient the abil-
 62

ity to communicate on that channel. It is perhaps best to think that a channel end-point
has been transmitted.
Let us consider a channel end-point that is transmitted across a domain boundary
over another channel that already crosses the boundary. If the transmitted end-point re-
mains functional, it provides a dynamically-established connection between the two
sides of the boundary. This is the kind of connection that firewalls typically forbid:
opening arbitrary network connections or allowing network-object proxy requests is not
allowed without further checks. The new channel that crosses the firewall could be seen
as an implicit firewall tunnel, but the establishment of trusted tunnels involves more
than simply passing a channel over another one, otherwise the firewall would loose all
control of security.
A firewall must watch the communication traffic over each channel that crosses it;
that is, it must act as an intermediary and forwarder of messages between the outside
and the inside of a domain. If a channel end-point is seen passing through, the firewall
must decide whether to allow communication on that channel, and if so it must create a
forwarder for it. So, a channel through a firewall must really be handled as two channels
connected by a filter [20].
Therefore, ability to communicate on a channel depends not only on possessing the
end-point of a channel, but also on where the other end-point of the channel is, and how
it got there. If the other end-point was sent through a firewall, then the ability to effec-
tively communicate on that channel depends on the attitude of the firewall.
Our approach: We provide a framework where processes exist in multiple disjoint
locations, and such that the location of a process influences its ability to communicate
with other processes. Dynamic connectivity is achieved by movement, but movement
does not guarantee continued connectivity.
Distribution
The jc-calculus has no inherent notion of distribution. Processes exist in a single contig-
uous location, by default, because there is no built-in notion of distinct locations and
their effects on processes. Interaction between processes is achieved by global, shared
names (the channel names); therefore every process is assumed to be within earshot of
every other process. Incidentally, this makes distributed implementation of the original
7t-calculus (and of CCS) quite hard, requiring some form of distributed consensus.
Various proposals have emerged to make the n-calculus suitable for distributed im-
plementation, and to extend it with location-awareness. The asynchronous 7t-calculus
[25] is obtained by a simple weakening of the 7t-calculus synchronization primitives.
Asynchronous messages simplify the requirements for distributed synchronization [32],
but they still do not localize the management of communication decisions. The join-cal-
culus [20] approaches this problem by rooting each channel at a particular process; this
provides a single place where synchronization is resolved. LLinda [18] is a formaliza-
tion of Linda [15] using process calculi techniques; as in distributed versions of Linda,
LLinda has multiple distributed tuple spaces, each with its local synchronization man-
ager.
Our approach: We restrict communication to happen within a single location, so
that communication can be locally managed. In particular, interaction is by shared lo-
cation, not by shared names. Remote communication is modeled by a combination of
mobility and local communication.
 63

Locality
By locality we mean here distribution-awareness: a process has some notion of the lo-
cation it occupies, in an absolute or relative sense.
A growing body of literature is concentrating on the idea of adding discrete loca-
tions to a process calculus and considering failure of those locations [4, 21]. A notion
of locations alone is not sufficient, since locations could all really be in the "same
place". However, in presence of failures one could observe that certain locations have
failed and others have not, and deduce that those locations are truly in different places,
otherwise they would all have failed at the same time. The distributed join-calculus
[21], for example, adds a notion of named locations, and a notion of distributed failure;
locations form a tree, and subtrees can migrate from one part of the tree to another,
therefore becoming subject to different failure patterns.
This failure-based approach aims to model traditional distributed environments,
and traditional algorithms that tolerate node failures. However, on the Internet, node
failure is almost irrelevant compared with inability to reach nodes. Web servers do not
often fail forever, but they frequently disappear from sight because of network or node
overload, and then they come back. Sometimes they come back in a different place, for
example, when a Web site changes its Internet Service Provider. Moreover, inability to
reach a Web site only implies that a certain path is unavailable; it implies neither failure
of that site nor global unreachability. In this sense, a perceived node failure cannot sim-
ply be associated with the node itself, but instead is a property of the whole network, a
property that changes over time.
Our approach: The notion of locality is induced not by failures, but by the need to
cross barriers. Barriers produce a non-trivial and dynamic topology of locations. Loca-
tions are observably distinct because it takes effort to move between them, and because
some locations may be "absent" and are distinguishable from locations that are
"present". Failure is only represented, in a weak but realistic sense, as becoming forever
unreachable.
Mobility
There are different senses in which people talk about "process mobility"; we try to dis-
tinguish between them.
A K-calculus channel is mobile in the sense that it can be transmitted over another
channel. Let's imagine a process as having a mass, and its channels as springs connect-
ing it to other processes. When springs are added and removed by communication, the
process mass is pulled in different directions. Therefore, the set of channels of a process
influences its location, and a change of channels causes the process to change location.
This is particularly clear if a process has a single active channel at any time, because a
single spring will strongly influence a process location. By this analogy, channel mo-
bility can be interpreted as causing process mobility.
However, our desired notion of process mobility is tied to the idea of crossing do-
main barriers. This is a very discrete, on-off, kind of mobility: either a process is inside
a domain, or it is not. Representing this kind of mobility by adding and removing chan-
nels is not immediate. For example, if a 7C-calculus channel crosses a barrier (that is, if
 64

it is communicated to a process meant to represent a barrier), there is still no clear sense

in which the process has also crossed the barrier. In fact, the same channel may cross
several disjoint domain barriers, but a process should not exist in all those domains at
once. Therefore, a 7i-calculus process can be made to span several domains, which is
something we would like to rule out from the start.
Our notion of process mobility is tied to the notion of nested domains. It still re-
mains to be seen how simple hierarchies, such as the ones we have emphasized in our
mental images, can capture the sometimes complex relationships between administra-
tive domains. Still, some notion of domain nesting seems natural, and this is difficult to
represent adequately in process calculi like the 7i-calculus that are based on "flat" sets
of processes.
Another, more direct, form of process mobility has been proposed: a process may
move by being transmitted over a channel to another process. This mechanism is not
present in the basic 7t-calculus, but is present in so-called higher-order 7i-calculi^. This
is certainly a form of process mobility of the kind found in mobile agent systems, where
"whole" and "alive" agents are transmitted over communication channels. However,
there are a couple of details that should raise some concern, again having to do with do-
main boundaries.
First, transmitting a process through a firewall over a channel relies on the exist-
ence of a channel that already crosses a firewall. That is, we still have to separately
model the establishment of firewall tunnels.
Second, if a process has connections before the move, does it maintain them on the
other side of the firewall? If the answer is "yes", then arbitrary channels can be made to
penetrate a firewall from the existence of a single channel. If the answer is "not neces-
sarily", why and how do those other channels break? If the firewall acts as a filter, then
it has to analyze the processes that are passing though and their communication capa-
bilities; this may be considerably more complex than analyzing simple messages.
Third, transmitting a process over a channel is a copy rather than move operation.
Nothing prevents the original process from continuing, and nothing prevents the trans-
mitted process from being further replicated. That is, the idea that the same process
moved from here to there seems to be lost.
Our approach: mobility is not represented by passing channels over channels, nor
by passing processes over channels. It is represented by processes jumping across
boundaries. The identity of the moving process is preserved because a process that
crosses a boundary disappears from its previous location.
Security
It is possible to extend a concurrent calculus with cryptographic primitives, as in the spi
calculus extension of the 7t-calculus [2]. Much fundamental progress is being made in
this direction.
In our approach, security is not tied to cryptographic primitives, but on the ability
or inability to cross barriers, which is conferred by capabilities. Given the mechanisms

^ The fact that the latter can be formally reduced to the former [34] is best ignored for
this discussion.
 65

already required to handle boundaries, the need for cryptographic extensions does not
arise immediately. For example, a boundary enclosing a piece of text can be seen as an
encryption of the text, in the sense that a capability, the cryptokey, is needed to cross
the boundary and read the text. There is an unexpected similarity between a firewall sur-
rounding a major company and the encryption of a piece of data, both being barrier-
based security mechanisms at vastly different scales.

3.2 Formalisms for Reactivity

An important part of interacting and computing on the Web is being able to react to
bandwidth fluctuations in real time. We will not talk about this subject further, prefer-
ring to concentrate on the other computational observables discussed in the Introduc-
tion. We just note that the vast literature on real-time reactive systems should be
relevant in this context (e.g., see [6]), and that some Web-related work in this area has
been carried out [12] and has found applications [27].

4 Ambients
We have now sufficiently discussed our design constraints and the deficiencies of ex-
isting solutions, and we are finally ready to explain our own proposal in detail. We want
to capture in an abstract way, notions of locality, of mobility, and of ability to cross bar-
riers. To this end, we focus on mobile computational ambients; that is, places where
computation happens and that are themselves mobile.
4.1 Overview
Briefly, an ambient, in the sense in which we are going to use this word, is a place that
is delimited by a boundary and where computation happens. Each ambient has a name,
a collection of local processes, and a collection of subamhients. Ambients can move in
and out of other ambients, subject to capabilities that are associated with ambient
names. Ambient names are unforgeable, this fact being the most basic security property.
In further detail, an ambient has the following main characteristics.
• An ambient is a bounded place where computation happens.
If we want to move computations easily we must be able to determine what parts
should move. A boundary determines what is inside and what is outside an ambient,
and therefore determines what moves. Examples of ambients, in this sense, are: a
web page (bounded by a file), a virtual address space (bounded by an addressing
range), a Unix file system (bounded within a physical volume), a single data object
(bounded by "self) and a laptop (bounded by its case and data ports). Non-examples
are: threads (where the boundary of what is "reachable" is difficult to determine) and
logically related collections of objects. We can already see that a boundary implies
some flexible addressing scheme that can denote entities across the boundary; exam-
ples are symbolic links, URLs (Uniform Resource Locators) and Remote Procedure
Call proxies. Flexible addressing is what enables, or at least facilitates, mobility. It
is also, of course, a cause of problems when the addressing links are "broken".
 66

• Ainbienls can be neslcd witliin other ainbicnls, forming a tree slruclure.

As we discussed, administrative domains are (often) organized iiierarcliically. Mo-
bility is represented as navigation across a hierarchy of ambients. For example, if we
want to move a running application from work to home, the application must be re-
moved from an enclosing (work) ambient and inserted in a different enclosing
(home) ambient.
• Each ambient has a collection of local running processes.
A local process of an ambient is one that is contained in the ambient but not in any
of its subambients. These "top level" local processes have direct control of the am-
bient, and in particular they can instruct the ambient to move. In contrast, the local
processes of a subambient have no direct control on the parent ambient: this helps
guaranteeing the integrity of the parent.
• Each ambient moves as a whole with all its subcomponents.
The activity of a single local process may, by causing movement of its parent, influ-
ence the location, and therefore the acdvity, of other local processes and subambi-
ents. For example, if we move a laptop and reconnect it to a different network, then
all the threads, address spaces, and file systems within it move accordingly and au-
tomatically, and have to cope with their new surrounding. Agent mobility is a special
case of ambient mobility, since agents are usually single-threaded. Agents, like am-
bients, automatically carry with them a collection of private data as they move.
• Each ambient has a name.
The name of an ambient is used to control access (entry, exit, communication, etc.).
In a realistic situation the true name of an ambient would be guarded very closely,
and only specific capabilities based on the name would be handed out. In our exam-
ples we are usually more liberal in the handling of names, for the sake of simplicity.
We have developed and are still studying a formal calculus of ambients [ 13]. In this
paper we stay away from formalism; the essential features of the ambient calculus will
be conveyed by (fairly precise) metaphors.
As a first metaphor, already partially outlined, we can consider a foreign travel sce-
nario. A country with border guards can be seen as an example of a static ambient. Local
officers (privileged processes) govern movement across the border and communication
within the border. Entry and exit capabilities (passports and visas) moderate mobility
across the border. Officers may decide who may join them as new officers.
As an example of a mobile ambient we may consider a tourist traveling to such a
country. The tourist travels by being transported by another mobile ambient: a train, or
airplane. The tourist's visas are checked on entry and exit. The tourist may be impris-
oned within the country (by withholding the exit capability) or, if not careful, even
killed. On the other hand, the tourist may be allowed to become a local officer (be nat-
uralized), and in such a function may act as a spy or a saboteur, sending out encrypted
messages or disabling subsystems.
In the following section we discuss a more abstract metaphor that is easy to de-
scribe and has an intuitive graphical presentation. It corresponds faithfully to the full,
formal, ambient calculus.
 67

4.2 The Folder Calculus: Mobility

In this office-style metaphor we represent an ambient as a folder. A folder confines its
contents: something is either inside or outside any given folder. Each folder has a name
that is written on the folder tag. Folders are naturally nested, and can be moved from
place to place.
The computational aspect of the calculus is represented by assuming that folders
are active. In addition to subfolders, folders may contain gremlins that cause the folder
to move around, and are moved together with their folder.

/T\

A folder with label n and contents P

We now describe the reductions (atomic computation steps) that can arise in the
folder calculus. Each reduction can happen at the "top level" of the computation (say,
"on the desktop") or inside a folder. That is, the folder hierarchy is active at every level,
with some caveats noted below.
The Enter Reduction
Our first reduction causes a folder to enter another folder. On the left of the arrow, in
the figure below, the folder labeled n contains a gremlin that is ready to execute the in-
struction in m (meaning: "let the surrounding folder enter a folder labeled m"), and then
continue with P. (The partial occlusion of P in the figure is meant to indicate that the
instruction in m must execute before P can be activated.) The collection of other grem-
lins and subfolders that may be contained in n is represented by the letter Q, and simi-
larly by R within m. A folder labeled m happens to exist near the folder n, where "near"
means "within the same surrounding folder, or on the desktop".

- ^

Enter reduction
In this situation the operation in m can execute, resulting in the configuration to the
right of the arrow. The result of the operation is that the folder n becomes a subfolder
of the folder m, and the gremlin who executed the instruction is ready to continue with
P. The instruction in m has been consumed. Any other gremlins or subfolders in Q and
R are unchanged.
A reduction can happen only if the conditions on the left of the arrow are satisfied.
That is, in m executes only if there is a sibling folder labeled m. Otherwise, the operation
remains blocked until a sibling folder labeled m appears nearby. Such a folder may ap-
pear, for example, because it moves near n, or because some of the gremlins in Q cause
/I to move near it.
 68

Many reductions can be simultaneously enabled, in which case one is chosen non-
deterministically. For example, there could be two distinct folders labeled m near n, in
which case n could enter either one. Also, there could be another gremlin, as part of Q,
trying to execute an in m operation, in which case exactly one of the gremlins would
succeed. Moreover, there could be another gremlin in Q trying to execute an inp oper-
ation, and there could be another folder labeled p near n, in which case the n folder could
enter either m or p.
The Exit Reduction
Our second reduction is essentially the inverse of the previous one: two nested folders
become siblings. On the left of the arrow, the folder labeled n contains a gremlin that is
ready to execute the instruction out m (meaning: "let the surrounding folder exit its par-
ent labeled m"), and then continue with P. The parent folder is in fact labeled m.
In this situation the operation out m can execute, resulting in the configuration to
the right of the arrow. The result of the operation is that the folder n becomes a sibling
of the folder m, and the gremlin who executed the instruction is ready to continue with
P. The instruction outm\\as been consumed. Any other gremlins or subfolders in Q and
R are unchanged.

JZI
r^ \i)ut my f • )
Q - ^
R

Exit reduction
The operation out m executes only if the parent folder is labeled m. Otherwise, the
operation remains blocked until a parent folder labeled m materializes. The parent may
become m, for example, because some of the gremlins in Q cause n to move inside a
folder labeled m, at which point out m can execute.
Again, several reductions may be enabled at the same time. For example, there
could be a gremlin in Q trying to execute an in p operation, and there could be a folder
labeled p in R. Then, the folder n could either exit m or enter p.
The Open Reduction
Our third reduction is used to discard a folder while keeping its contents. In the picture
it is helpful to imagine that there is a folder surrounding the entities on the left of the
arrow, so that open n followed by P is a gremlin of that folder, and n is one of its sub-
folders. The gremlin is ready to execute the instruction open n (meaning: "let a nearby
folder labeled n be opened"), and then continue with P. A folder labeled n happens to
exist nearby.
In this situation the operation open n can execute, resulting in the configuration to
the right of the arrow. The result of the operation is that the folder n is discarded, but its
contents Q are spilled in the place where the folder n used to be. The instruction open n
has been consumed, and the gremlin is ready to continue with P.
 69

I open
SD a -^ Open reduction
As before, the operation open n is blocked if there are no folders labeled n nearby.
If there are several ones, any one of them may be opened.
The Copy Reduction
Our fourth reduction is used to replicate folders and all their contents. On the left of the
arrow, a copy machine is ready to make copies of a folder P (actually, P could also be
a gremlin or any collection of folders and gremlins). We should imagine that the origi-
nal P is firmly placed under the cover of the copy machine, and is compressed into im-
mobility: none of the gremlins or folders of P can operate or move around (otherwise
the copy might be "blurred"). However, as soon as a copy of P is made, that copy is free
to execute.

P P
T ^ —^ T ^ p
Copy reduction
The copy machine can produce a new copy of P and of all its contents at will (no-
body needs to push the copy button). After that, the copy machine can operate again,
indefinitely. So, on the right of the arrow we have the same configuration as on the left,
plus a fresh copy of P. We could think that copies of/" are made on demand, whenever
needed, rather than being continuously produced.
Name Creation
The handling of names is a delicate and fundamental part of our calculus. Fortunately,
it is very well understood: it comes directly from the n-calculus.
Name creation is an implicit operation, in the sense that there are no reductions as-
sociated with it. It is represented below as the creation of a rubber stamp for a name n,
which can be used to stamp folder labels. Any number of folders can be stamped with
the same rubber stamp.

-S-
: ^ ;
"> X

A rubber stamp is used not as much to give a name, as to give authenticity to a fold-
er. There are several components to this notion, and we need to stretch our office met-
aphor a bit to make it fit with the intended semantics.
At a microscopic scale, each rubber stamp has slight imperfections that can be used
to tell which rubber stamp was used to stamp each particular folder. Therefore, the par-
ticular name chosen for a rubber stamp is irrelevant: what is really important is the re-
 70

lationship between a rubber stamp and the folders it has stamped. Humans, however,
like to read names, not microscopic imperfection, so we keep names associated with
rubber stamps and folder labels. Still, we are free to change those names at any time, as
long as this is done consistently for a rubber stamp, all its related folders, and all the
other uses of the name on the rubber stamp. This consistent renaming could be consid-
ered as a reduction, but it is simple enough to be considered as a basic equivalence be-
tween configurations of folders, expressing the fact that superficial names are not really
important.
In our metaphor, a copy machine can be used to copy anything contained in a fold-
er, including rubber stamps. Therefore, even if we started with all the rubber stamps
having different names, eventually there might be multiple rubber stamps carrying the
same name. To make authenticity work, we have to assume that copy machines cannot
copy rubber stamps perfectly at the microscopic level: when a rubber stamp is replicat-
ed, a different set of microscopic imperfections is generated. That is, rubber stamps are
unforgeable by assumption.
For all these reasons, two rubber stamps carrying the same name n are really two
different rubber stamps. To preserve authenticity, we do not want these rubber stamps,
and the folders they stamp, to get confused. In our visual representation, we collect all
the folders stamped by a rubber stamp, and all the other occurrences of its name, within
a dashed boundary: this way we can always tell, graphically, which folders were authen-
ticated by a rubber stamp, even when different rubber stamps have the same name.
This dashed border is a flexible boundary and can move about fairly freely (it is just
a bookkeeping device). We have three main invariants for where a dashed border can
be placed. First it must always be connected with its original rubber stamp. Second, it
must always enclose all the folders that have ever been stamped with its particular
stamp and all other occurrences of the name (e.g. within gremlin code); if a folder
moves away, the dashed boundary may have to be enlarged. Third, dashed boundaries
for two rubber stamps with the same name must not intersect; if we should ever need to
do so we shall first systematically rename one of the two rubber stamps and the related
names, so that there is no confusion. The dashed boundaries for rubber stamps with dif-
ferent names can freely intersect.

allowed forbidden
It is allowable to nest dashed boundaries for rubber stamps with the same name: an
occurrence of that name will refer to the closest enclosing rubber stamp, in standard
block scoping style.
Leaves of the Syntax
At the bottom of our syntax there is the inactive gremlin, which can be represented as
an empty border. The inactive gremlin has no reductions.
 71

o
Inactive gremlin
Inactive gremlins are often simply discarded or omitted. For example, multiple in-
active gremlins can be collapsed into one inactive gremlin, and a folder containing only
one inactive gremlin is usually represented as an empty folder.

0 0 = 0

£3
Programs in the folder calculus are built from these foundations, by assembling
collections of gremlins, folders, rubber stamps, and copy machines, and possibly plac-
ing them inside other folders.
The Theoretical Power of Mobility
Before moving on to our next and final reduction, we pause and consider the operations
we have introduced so far. These operations are purely combinatorial, that is, they in-
troduce no notion of parameter passing, or message exchange. Also, they deal purely
with mobility, not with communication or computation of any familiar kind. Yet, they
are computationally complete: Turing machines can be encoded in a fairly direct way
(see [13]).
Moreover, very informally, it is possible to see an analogy between the Enter re-
duction and increment, the Exit reduction and decrement, the Open reduction and test,
and the Copy reduction and iteration. Therefore all the ingredients for performing arith-
metic are present, and it is in fact possible to represent numbers as towers of nested fold-
ers, with appropriate operations.
Data structures such as records and trees can be represented easily, by nested fold-
ers; folder names represent pointers into such data structures. Storage cells can be rep-
resented as folders whose contents change in response interactions with other folders;
in this case a folder name represents the address of a cell.
In summary, the mobility part of the Folder Calculus already has the full power of
any computationally complete language. In principle, one could use it as a graphical
scripting language for the office/desktop metaphor.

4.3 The Folder Calculus: Communication

Mobility is by itself a computationally complete paradigm. Still, we want to talk about
communication in a direct way; for example, gremlins may want to exchange names of
folders to visit. Representing communication within the existing set of mobility primi-
tives is theoretically possible, but unbearably clumsy. So, our next task is to introduce
communication primitives that do not invalidate the principles described in Section 3,
and that do not spoil the folder metaphor. The primitives we introduce next are probably
 72

the simplest imaginable. In particular, they do not conflict with the notion of strict fold-
er containment, and do not duplicate mobility functionality.
The Read Reduction
We begin by introducing a new entity that can sit inside a folder: a message. To remain
within the office metaphor, we imagine writing messages onto throw-away Post-it notes
that are attached to the inside of folders.
A gremlin can write the name of a folder on a note, and can attach the note to the
current folder (the folder the gremlin is in). This is an output operation, and is represent-
ed graphically by a message written on a note. We shall discuss shortly what are the par-
ticular messages used in the folder calculus. More generally, we may imagine writing
any kind of data as a message; in this view, a note can be seen as nameless data file that
is kept within a folder.

1\
M

Output
Conversely, a gremlin can grab any note attached to the current folder, read its mes-
sage, discard the note, and proceed with the knowledge of the message. This is an input
operation, and is represented graphically by a process P with occurrences of the variable
X (written P[x]) that is waiting for a note with a message to be bound to x.

Input
The Read reduction is the interaction between input and output operations or,
equivalently, between message notes and input operations. In a situation where an input
and an output are present, the Read reduction can execute, resulting in the configuration
on the right of the arrow, which is simply P{ A/}: the residual gremlin P that has read M
into X. The note that appears on the left of the arrow is discarded: it is consumed by read-
ing and does not appear on the right. (If the note needs to persist, it can be replicated by
a copy machine.)

T^
M —^ P{M]

Read reduction
An input operation blocks if there are no available messages. Several input opera-
tions may contend for the same message and only one of them will obtain it and be able
to proceed. An output operation, however, never blocks (it is asynchronous); it simply
drops a message in the current folder and has no continuation.
Inputs and outputs usually happen within a folder (although the reduction above al-
 73

lows them to happen on the desktop). Within such a folder, the identity of the input or
output gremlins are not important: anybody can talk to anybody else, in the style of a
chat room. In practice, different folders will be dedicated to different kinds of conver-
sation, so one can make assumptions about what should be said and what can be heard
in a given folder. This idea gives rise to a type system for the folder calculus [14].
Messages and Capabilities
We can imagine many kinds of messages that can be written on Post-it notes. Here we
focus on messages that are capabilities: they allow the reader of a message to perform
privileged actions. There are two kinds of capabilities, used in different contexts: nam-
ing capabilities and navigation capabilities.
Naming capabilities are simply ambient names used as messages. A name n can be
seen as a capability to construct (and rubber-stamp) a folder named n.
We have already seen the main navigation capabilities, implicitly. We have pre-
sented the operations in n, out n and open n as three distinct operations followed by a
continuation P. In fact, they are special cases of a single operation M.P, where A/ is a
navigation capability and P is the continuation after navigation. Given a name n, in n is
the capability to enter an ambient named n, out n is the capability to exit an ambient
named n, and open n is the capability to open an ambient named n. Navigation capabil-
ities are extracted from naming capabilities, meaning that knowing n implies the ability
to construct, for example, in n, but knowing in n does not imply knowing n.
Navigation capabilities can be composed into navigation paths. For example, in n.
out m. open p. out ^ is a path capability that can be written in a single message, read into
an input variable x, and executed in its entirety by x.P (assuming, of course, that the path
can be followed). It useful to have an empty navigation path, written here, such that
here.P has no effect and continues with P, and M.here = here.M = M.
Example
This is an artificial example that uses each of the five folder reductions once. We use a
single dashed line for multiple rubber stamps when the order of nesting of dashed lines
does not matter. We remove rubber stamps when they are no longer needed.

P1 \ n l I m I

Enter
 74

^—————
/
rr\
r?ir^\
Open ' ppi Copy

^1 \outp\
• t^

Read Exit

/T"V £1^

The reader is encouraged to follow the reductions, comparing them with their defini-
tions.
Example: Adobe Distiller
Adobe Distiller is a program that converts files in Postscript format to files in Adobe
Acrobat format. The program can be set up to work automatically on files that are
placed in a special location. In particular, when a user drops a Postscript file into an in-
box folder, the file is converted to Acrobat format and dropped into a nearby outbox
folder.
The following figure describes such a behavior. The distiller folder contains the in-
box and outbox folders mentioned above; outbox is initially empty. The input folder
contains the file the user wants to convert, in the form of a message. The input folder
contains also a gremlin that moves the input folder into the inbox. (We can imagine that
this piece of gremlin code is generated automatically as a result of the user dragging the
input folder into the inbox folder.)
 75

X distiller\
/outbo^^

open input
S distill(xj\
I out inhox. in i>ulhoJ^>\

%!PS.^ti,. I in distiller, in mfc77?p>]

The inbox contains the program necessary to do the format conversion and drop the
resuh into the outbox. First, any input folder arriving into the inbox must be opened to
reveal the Postscript file; this is done by the copy machine on the left. Then, any such
file is read; this is done by the copy machine on the right. As a result of each read, an
output folder is created to contain a result. Inside each output folder, a file is distilled
(by the external operation distill{x)) and left there as an output. The output folder is
moved into the outbox folder.
It should be noted that the program above represents highly concurrent behavior,
according to the reduction semantics of the folder calculus. Multiple files can be
dropped into the inbox and can be processed concurrently. The opening of the input
folders and the reading of their contents is done in a producer-consumer style. More-
over, each distilling process may be executing while its output folder is traveling to the
outbox. Representing this behavior in an ordinary concurrent language would not be en-
tirely trivial; here we have been able to express it without cumbersome locking and syn-
chronization instructions.
Example: Synchronous Output
It is sometimes useful to know that a message has been read by somebody before pro-
ceeding. Synchronous output is an output operation with a continuation that is triggered
only after the output message has been read.
Synchronous output is expressible within the folder calculus, if we assume that the
calculus has been extended to allow the exchange of pairs of values (this extension can
in fact be encoded within the calculus). Together with synchronous output, we need to
define a matching input operation. These new operations are depicted with additional
striped borders in the figures.
The synchronous output of a message M is obtained by creating a fresh name, k,
outputting (asynchronously) the pair M,k, and waiting for the appearance of a folder
named k before proceeding.
 76

.o_]-
tl^ - ;^@ niF^]
The corresponding input operation for a variable x, is obtained by expecting an input
pairx,^, creating an empty folder named k (which triggers the synchronous output con-
tinuation), and continuing with the normal use of the input jc.

MJ
Therefore, when the process P starts running, it can assume that somebody has read the
message M.
4.4 Security
The folder calculus has built-in features that allow it to represent security and encryp-
tion situations rather directly; that is, without extending the calculus with ad-hoc prim-
itives for encryption and decryption. In this section, we discuss a number of examples
based on simple security protocols.
We should make clear here what we mean by security for the folder calculus. Se-
curity problems arise at every level of a software system, not just at the cryptographic
level. Given any set of security primitives, and any system written with those primi-
tives, one can ask whether the system can be attached at the "low-level", by attacking
weaknesses in the implementation of the primitives, or at the "high-level", by attaching
weaknesses in how the primitives are used. Efforts are underway to study the security
of high-level abstractions under low-level attacks [1], but here we are only concerned
with high-level attacks. That is, we assume that an attacker has at its disposal only the
primitives of the folder calculus. This is the kind of attack that a malicious party could
mount against honest folders interacting within a trusted server or over a trusted net-
work. For example, even within a perfectly trusted server, if a folder gives away its
name a, it could be killed by an attacker performing open a.
Authentication
In this example, a home folder is willing to let any folder in, but is willing to open only
those folders that are recognized as having originally come from home. Opening a fold-
er implies conferring top-level execution privilege to its gremlins, and this privilege
should not be given to just anybody. A particular home gremlin that has execution priv-
ilege wants to leave home and then wants to come back home and be given the same
execution privilege it enjoyed before.
The mechanism the home folder uses to recognize the gremlin is a pass: a use-once
authentication token. Passes are generated in the left-hand side of the figure below,
where a copy machine produces fresh configurations, each consisting of a new rubber
 77

Stamp, a single message (the pass) stamped by the rubber stamp, and a single open ca-
pability for the name on the rubber stamp.

The traveling gremlin is on the right-hand side of the figure. It inputs a pass n by
reading a message into a variable x, and it eventually uses the pass to label a folder. The
gremlin, in the form of the folder g, takes a short walk outside and comes back. Then it
exposes a folder named n, which is opened by the corresponding open n capability that
was left behind. The gremlin P can then continue execution at the top level of home; for
example, it may read another pass and leave again.
Since the scope of each pass n is restricted by a locally-generated rubber stamp, the
capability open n is not at risk of ever opening some foreign folder. There are actually
two underlying security assumptions here. The first is that nobody can accidentally or
maliciously create a pass that matches n: this can be guaranteed in practice with arbi-
trarily high probability. The other assumption is that nobody can steal the name n from
the traveling gremlin. This seems very hard or impossible to guarantee in general, par-
ticularly if the gremlin visits a hostile location that disassembles the gremlin by low-
level mechanisms (below the abstraction level of the folder calculus). However, if the
gremlin visits only trusted locations through trusted networks (ones that preserve the
abstractions of the folder calculus), then no interaction can cause the gremlin to be un-
willingly deprived of its pass.
Nonces
A nonce is a use-once token that can be used to ensure the freshness of a message. In
the example below, a nonce is represented by a fresh name n. The folder a sends the
nonce to the folder b, where the nonce is paired with a message M and sent back.
The folder a then checks that the nonce returned by h is the same one that was sent
to b. This is achieved by creating an empty folder named by the returned nonce, and try-
ing to open that folder with the original nonce. If the test is successful, then a knows
that the message M is fresh: it was generated after the creation of n.
If nonce and msg are public names, then an attacker could disrupt this protocol by
destroying the message folders in transit. Even worse, an attacker could inject nonce
and msg folders into aor b containing misleading information or damaging code. If a
and b have already established shared keys, they can avoid these problems by exchang-
ing and opening only messages encrypted under the keys (this is discussed shortly).
 78

m. <ipen rumce ^ "^

n' Imsg
I out h. in a ^ ^ ^

^ ^

An attacker could also impersonate a orfeby creating folders with those names, and
could then intercept messages. However, the names of principals like a and b will nor-
mally be closely guarded secrets, so that impersonation cannot happen. In contrast, ca-
pabilities like in a will be given out freely, since the act of entering a folder cannot by
itself cause damage to the folder, even if who enters is malicious.
Shared Keys
A name can be used as a shared key, as long as it is kept secret and shared only by cer-
tain parties. A shared key can be reused multiple times, e.g., to encrypt a stream of mes-
sages.
A message encrypted under a key k can be represented as a folder that contains the
message and whose label is k. We call such a folder a k-en\dope for the message.
Knowledge of k (or, at least, of the capability open k) is required to open the folder and
read the message.

& generation of a fresh shared key k

£T\ encryption:
~M\ plaintext M inside a A:-envelope

I open k decryption:
opening a ^-envelope and reading the contents
\

To continue the Authentication example, a traveling gremlin could carry a shared

key k, generated inside the home folder, and send messages back home inside ^-enve-
lopes. The home folder could decrypt those messages by using the shared key. If the
shared key is unique to a particular gremlin, this has the effect of authenticating the
source of the messages.
In many classical distributed protocols, shared keys are assume to be distributed by
unspecified covert means, separately from ordinary communication on public channels.
The principals are assumed to have already established shared keys before they begin
 79

interacting. The icey distribution problems is left as a separate problem.

In our framework, which includes mobility, some aspects of key distribution can
be modeled explicitly. For example, a key can become shared between distant princi-
pals because they generate the key when they are in the same location, and then move
apart. Or, a courier folder can transport a key from a principal to another over a trusted
network, and then the principals can use the key to communicate over an untrusted net-
work.

4.5 Textual Syntax

So far we have been relying on a visual syntax, but we are going to need some textual
syntax for writing further examples. The following table summarizes the folder calculus
visual syntax, and introduces the corresponding textual syntax. There is a one-to-one
correspondence between textual syntax and visual syntax; therefore, it is possible to
freely mix them, if desired, nesting either one inside the other.
Correspondence Between Textual and Visual Syntax
We use P,Q,R to range over ambient and process expressions, and M,N to range over
message expressions. As shown in the table, the creation of a new name is written {vn)P
where the Greek letter v (nu) binds the name n within the scope P. An ambient is written
n[P\ where n is the ambient name, where the brackets denote the ambient boundary, and
where P is the contents of the ambient.
Processes P,Q,R
Textual Syntax Visual Syntax Comments

, New name /r in a scope P.

f
V ...

rr\ Folder (ambient) of name n and contents

n{P]
p P.

M.P [M^T^ Action M followed by P.

Two processes in parallel.

P\Q P Q (Visually: contiguously placed in 2D.)

0 o Inactive process (often omitted).

p
\P Replication of P.
 80

(M) M Output M.

{n).P
\D Input n followed by P.

(P)
O Grouping.

Messages M,N

n A name
in n An entry capability
out n An exit capability
open n An open capability
here The empty path of capabilities
M.N The concatenation of two paths

The textual syntax for the folder calculus is in fact the full syntax of the ambient
calculus we were alluding to previously. Therefore, our folder calculus metaphor is
quite exact: the syntax and semantics of our formal ambient calculus [ 13] has now been
completely explained through the metaphor.
Example: Adobe Distiller
This is the textual representation of the example in Section 4.3.
distiller{
inbox[
\open input I
!(x) output[{distill{x)) I out inbox. in outbax]] I
outhox[\]

input[{"%JPS...") I in distiller, in inbox]

5 Ideas for Wide-Area Languages

The ambient calculus is a minimal formalism designed for theoretical study. Our final
goal, though, is to program the Internet; for this we are certainly going to need some-
thing more elaborate and convenient than a formal calculus. Still, the basic constructs
of the ambient calculus represent our understanding of the fundamental properties of
mobile computation over wide-area networks. Therefore, we aim to find programming
 81

constructs that are semantically compatible with the principles of the ambient calculus,
and consequently of wide-area-networks.
These principles include (A) WAN-soundness: a wide-area network language can-
not adopt primitives that entail action-at-a-distance, continued connectivity, or security
bypasses, and (B) WAN-completeness: a wide area network language must be able to
express the behavior of web surfers and of mobile agents and users.
5.1 Related Languages
Many software systems have explored and are exploring notions of mobility and wide-
area computation. Among these are:
• Obliq [11]. The Obliq language attacks the problems of remote execution and
mobility for distributed computing. It was designed in the context of local area
networks. Within its scope, Obliq works quite well, but is not really suitable for
computation and mobility over the Web, just like most other distributed para-
digms developed in pre-Web days.
• Telescript [37]. Our ambient model is partially inspired by Telescript, but is al-
most dual to it. In Telescript, agents move whereas places stay put. Ambients,
instead, move whereas processes are confined to ambients. A Telescript agent,
however, is itself a little ambient, since it contains a "suitcase" of data. Some
nesting of places is allowed in Telescript.
• Java [23]. Java provides a working paradigm for mobile computation, as well as
a huge amount of available and expected infrastructure on which to base more
ambitious mobility efforts. The original Java mobility model, however was
based on mobility of code, not mobility of data or active computations. Data mo-
bility has now been achieved by the Java RMI extension, but computation mo-
bility (e.g. for threads or live objects) is still problematic.
• Linda [15]. Linda is a "coordination language" where multiple processes inter-
act in a common space (called a tuple space) by dropping and picking up tokens
asynchronously. Distributed versions of Linda exist that use multiple tuple spac-
es and allow remote operations over those. A dialect of Linda [16] allows nested
tuple spaces, but not mobility of the tuple spaces.
• The Join Calculus Language [22] is an implementation of the distributed Join
Calculus. The plain Join Calculus introduced an original and elegant synchroni-
zation mechanism, where a procedure invocation may be triggered by the join of
multiple partial invocations originating from different processes. The Distribut-
ed Join Calculus extends the Join Calculus with an explicit hierarchy of loca-
tions. As we already mentioned, the nature of this calculus makes distributed
implementation relatively easy. Migration of locations is allowed within the hi-
erarchy. Behavior of the system under a failure model is being investigated.
• WebL [27] is a language that specializes on fetching and processing Web pages.
It uses service combinators [12] to retrieve streams of data from unreliable serv-
ers, and it uses a sophisticated pattern matching sublanguage for analyzing struc-
tured (but highly variable) data and reassembling it.
 82

5.2 Wide-Area Languages

We do not have a full-blown language or library to propose yet; what is more important
here is the general flavor of such a language or library. In this section we discuss pro-
gramming constructs that are directly inspired by the ambient calculus and that are not
usually found in standard programming environments. In the remaining sections we
show a detailed example involving some of these constructs, and we speculate on a spe-
cific wide-area application that could hopefully be better written using a wide-area lan-
guage.
Ambients as a Programming Abstraction
Our basic abstraction is that of mobile computational ambients. The ambient calculus
brings this abstraction to an extreme, by representing everything in terms of ambients
at a very fine grain. In practice, to be useful as a programming abstraction, ambients
would have to be medium or large-grained entities. Ambient contents should include
standard programming subsystems such as modules, classes, objects, and threads.
In that sense, ambients could be used simply as wrappers of ordinary software sub-
systems, for the purpose of rendering those subsystems mobile. An ambient library
could include mechanisms for creating empty ambients, adding threads, objects, and
subambients to them, allowing threads to communicate with other ambients, and, of
course, allowing ambients to move around.
We should notice that the ability to smoothly move a collection of running threads
is almost unheard of in current software infrastructures. In this sense, ambients would
be a novel and non-trivial addition to our collection of programming abstractions.
Names vs. Pointers
The only way to use an ambient is by its name. No matter how we organize our hierar-
chies of ambients, the only way to manipulate the resulting structure is by using ambient
names. These names are detached from their corresponding ambients; one may possess
a name without having immediate access to any ambient of that name.
Therefore, names are like pointers, providing access to structures, but are only
"symbolic pointers": more like file names and URLs than memory addresses. Like file
names, these pointers need not always denote the same structure: they denote any am-
bient of the corresponding name, and this may change over time. When these pointers
denote no ambient, they are not "broken" (like dangling pointers, or illegal file names)
but rather "blocked" until a suitable ambient becomes available. Of course, one might
feel nervous about the possibility of blocking whenever one tries to use a name. On the
other hand, a blocked computation can be unblocked by providing an appropriate am-
bient, thus modeling both dynamic linking and the installation of plug-ins.
In an ambient-based language, every pointer to a data structure or other resource
outside of a given ambient should behave like a name. This is necessary to allow ambi-
ents to move around freely without being restrained by immobile ties.
Locations
Ambients can be used to model both physical and virtual locations. Some physical lo-
cations are mobile (such as airplanes) while others are immobile (such as buildings).
Similarly, some virtual locations are mobile (such as agents) while others are immobile
 83

(such as mainframe computers). These mobility distinctions are not reflected in the se-
mantics of ambients, but can be added as a refinement of the basic model, or embedded
in type systems that restrict the mobility of certain ambients.
Migration and Transportation
Each ambient is completely self-contained, and can be moved at any time with all its
running computations. If an ambient encloses a whole application, then the whole run-
ning application can be moved without need to restart it or reinitialize it. In practice, an
application will have ties to the local window system, the local file system, etc. These
ties, however, should only be via ambient names. When moving the applications, the
old window system ambient, say, becomes unavailable, and eventually the new window
system ambient becomes available. Therefore, the whole application can smoothly
move and reconnect its bindings to the new local environment. Some care will still be
needed to restart in a good state (say, to refresh the application window), but this is a
minor adjustment compared to what one would have to do if hard connections existed
between the application and the environment [5].
Communication
The basic communication primitives of the ambient calculus are based on the asynchro-
nous model and do not support global consensus or failure detection. These properties
should be preserved by any higher-level communication primitives that may be added
to the basic model, so that the intended semantics of communication over a wide-area
networks is preserved.
The ambient calculus directly supports only local communication within an ambi-
ent. Remote communication (for example, RPC) can be interpreted as mobile output
packets that transport and deposit messages to remote locations, wait for local input, and
then come back. The originator of an RPC call may block for the answer to come back
before proceeding, in the style of a synchronous call. In this interpretation, the outgoing
and incoming packets may get stuck for an arbitrary amount of time, or get lost. There
may be no indication that the communication has failed, and therefore the invoking pro-
cess may block forever without receiving a communication exception. This is as it
should be, since arbitrary delays are indistinguishable from failures. (Note, though, that
a time-out mechanism is easily implemented, by placing a remote invocation in parallel
with another activity that waits a certain time and, if the invocation has not completed,
cause something else to happen.)
Other examples of derived communication mechanisms include parent-child com-
munication, and communication between siblings (perhaps aided by the common par-
ent). All these appear quite useful, and will likely need to be included in any convenient
language.
Data Structures
Basic data structures, such as booleans and integers, can be encoded in terms of ambi-
ents, but the encodings are not practical. Therefore, basic types should be taken as prim-
itive, as usual.
Ambients can directly express hierarchies, so it should not be surprising that they
can easily represent structured data types. For example, a record structure of the form
 84

{li=ei,..., l„=en] can be represented as:

r[l\[]({e\)\open accessor)] I ... I ln[\{(e„)\open accessor)]]
where the name r is the address of the record, and the record fields /,=e, are represented
by subambients of the record ambient. The values e, are placed into the subambients as
outputs, waiting for an accessor ambient to go inside and read them.
The ambient semantics, though, is a bit richer than the minimum required for data
structures. For example, we would not normally want to allow the fields of a record to
take off and leave. Some possibilities supported by the ambient semantics are intriguing
but probably not worth the trouble, such as the ability to concatenate records by opening
them inside a new record.
Therefore, it would be prudent to build-in ordinary data structures, without relying
too much on the possibilities provided by the ambient semantics. Still, for ambients to
work at all we need these data structures to be mobile. Therefore, the "address" of a data
structure should be a proper ambient name, in the sense that it should provide blocking
behavior on access if needed.
Recursion
Many behaviors are naturally defined recursively. It is convenient to use a recursion
construct of the form, for example, rec X. P, representing a recursively defined process
P where any occurrence of X again denotes P. More generally, one can introduce mu-
tually recursive definitions, as done below when discussing resources.
The ambient calculus embeds what is essentially an iteration construct, \P. In terms
of theoretical power, this is sufficient to encode recursion. As a programming construct,
though it is inconvenient, and is also hard to implement because of its flavor of un-
bounded generation. So, for programming practice !P should be removed in favor of re-
cursion facilities.
Synchronimtion
Synchronization primitives are needed to coordinate the activities of multiple processes
within an ambient. In the ambient calculus, it is easy to represent basic synchronization
constructs, such as mutexes:
release n\P = n[]\P release a mutex called n, and do P
acquire n;P 4 open n. P acquire a mutex called n, then do P
where the open instruction blocks until a mutex is released, and consumes the mutex.
A useful technique is to synchronize on the change of name of an ambient. That is,
a process may wait for an ambient to change its name to a given name, and another pro-
cess may perform the change to trigger the first process. Name change can be represent-
ed in the ambient calculus, but not in a way that is atomic with respect to arbitrary
actions (particularly, movement) performed by other processes. A proper solution
would require all the processes in an ambient to synchronize on a mutex before any
movement or name change. This is a bit laborious; therefore it is convenient to intro-
duce name change as a primitive with the following atomic reduction:
n[be m.P I Q] -> m{P I Q]
 85

Static and Dynamic Binding

Wide-area systems are outside the control of any single administrator, and therefore
their maintenance cannot rely entirely on static configurations. In programming lan-
guage terms, they cannot rely completely on static binding of variables. A programer
who needs to change a static binding can usually change it, recompile the program, and
restart it. But one cannot easily restart the Internet, or any large system deployed across
it. Therefore some form of dynamic binding is necessary.
The names of the ambient calculus represent an unusual combination of static and
dynamic binding. Formally, the names obey the classical rules of static scoping, includ-
ing consistent renaming, capture-avoidance, and block nesting. However, the naviga-
tion primitives behave by dynamically binding a name to any ambient that has the right
name. The connection between a name and an ambient of that name is fully dynamic,
much like in dynamic linking.
The basic ambient calculus does not have any definitional facilities, but these are
clearly needed when building any large-scale systems, or even medium scale examples.
We are going to need definitional facilities for both static and dynamic binding, corre-
sponding to the underlying semantic features.
Statically-bound definitions can be represented by fairly normal let constructions.
A process-producing definition can be written:
letf{x\...Xn) = P; P' (\tif{x\...Xn) be P within the scope P')
with a standard expansion semantics within the scope P'. For call-by-value parameter
passing, assuming that V]...V„ are fully-evaluated messages, we have;
f{V,...V„) 4 p{x,^V,...Xn^Vn)
Dynamically bound definitions are much less routine, and more closely tied to the
ambient semantics; we discuss them next.
Resources
In the mobile agent paradigm, sites differ from each other by the collection of resources
that they provide to visiting agents. Agents bind dynamically to these resources as they
enter the ambient. Similarly, it is natural to think of each ambient as providing a collec-
tion of local resources, which could be seen as forming the interface, namespace, envi-
ronment, method suite, etc., of the ambient.
To this end, we introduce a mechanism of local resource definitions that is compat-
ible with the general ambient semantics, and can be easily explained in terms of the
primitives of Section 4. These resource definitions may occur only at the top level of an
ambient, and have the following form:

defj{x]...x„) - P\ (bind the resource/(.;ci...x„) to P within the ambient)

Note that the name/is not bound by this definition; it must be bound by a v binder at
some sufficiently global level. This way, different resources named/can be provided
in different ambients, and each use of the resource/will be relative to a given ambient.
A typical such resource could be the local window-system library.
A resource/can be invoked as follows, within an ambient where it is defined:
 86

f{y\...V„)lhere k P{x,^V|...x„^V„)
where the suffix "/here" indicates that a definition of/has to be found in the ambient
where/is invoked. In general, a path may be used in place of here, in which case the
definition of the resource is retrieved from the ambient obtained by following the path,
and the invocation is executed there. In other words, the effect of/(V|... V„)//7 is to trans-
port the invocation/(Vi...V„) along the path p, and then to invoke it within the target am-
bient. This can be expressed by two rules that each transport an invocation one step
further:
fiV,...V„yinn.p\n[Q] 4 n[Q\J{V,...V„)/p]
n\fiVi...V„)/out n.p I Q] 4 f(V,...V„)/p I n[Q]
For example, we could have an expression such as the following:
n[defxQ = (l);
deffiy) = x{)lhere\(x').{x'+y)]
\fO)linn
producing, after the invocation/(3)/in n:
n{defxQ = (\);
deffiy)=xO/here I {x'). {x'+y);
(4)]
The definition and invocation of resources can be encoded in the pure ambient cal-
culus; this way, resources automatically acquire an ambient-like flavor. In particular, if
an ambient has no definition for an invocation, the invocation blocks until the definition
becomes available in the designated place. If an ambient has multiple definitions for an
invocation, any one of them may be used. If an ambient containing definitions is
opened, its definitions become definitions of its parent.
ModuUuization
An ambient P that includes a collection of local definitions can be seen as a module, or
a class. More precisely, [P can be seen as a module or a class (since P there is inactive),
while any active P generated from it can be seen as a module instance or an object.
The action of performing an open on such an ambient can be seen as importing
from a module or inheriting from a class, since ambient definitions are transplanted
from one ambient to another. Moreover, one can regard the notation/(Af)/p either in
module-oriented style as p.j{M) (the invocation oif{M) from the module at/?), or in ob-
ject-oriented style as delegation [36] of /(A/) to the object found at p.
When seen as modules or components, ambients have several interesting and un-
usual properties.
First, ambients ^XQfirstclass modules, in the sense that one can choose at run time
which particular instance of a module to use.
Second, ambients support dynamic linking: missing subsystems can be added to a
running system simply by placing them in the right spot.
Third, ambients support dynamic reconfiguration. In most module and class sys-
 87

terns, the identity of individual modules is lost immediately after static or dynamic link-
ing. Ambients, though, maintain their identity at run time. As a consequence, a system
composed of ambients can be reconfigured by dynamically replacing an ambient with
another one. The blocking semantics of ambient interactions allows the system to
smoothly suspend during a configuration transition. Moreover, the hierarchical nature
of ambients allows the modular reconfiguration of entire subsystem, not just of individ-
ual modules. Dynamic reconfiguration is particularly valuable in long-running and
widely-deployed systems that cannot be easily stopped (for example, in telephone
switches); it is certainly no accident that thinking about the Internet led us to this prop-
erty.
Therefore, ambients can be seen as proper software components, according to a
paradigm often advocated in software engineering, where the components are not only
replaceable but also mobile.
Security
Ambient security is based on capabilities and on the notion that security checks are per-
formed at ambient boundaries, after which processes are free to execute until they need
to cross another boundary. This is a capability-based model of security, as opposed to a
cryptography-based model, or an access-control based model.
These three models are all interdefinable. In our case, access control is obtained by
using ambients to implement RPC-like invocations that have to cross boundaries and
authenticate every time. The cryptographic model is obtained by interpreting encryp-
tion keys as ambient names, which are by assumption unforgeable. Then, encryption is
given by wrapping some content in an ambient of a given name, and decryption is ob-
tained by either entering or opening such an ambient given an appropriate capability
(the decryption key).
Summary
We believe we have sufficiently illustrated how the ambient semantics naturally induc-
es unusual programming constructs that are well-suited for wide-area computation. The
combination of mobility, security, communication, and dynamic binding issues has not
been widely explored yet at the language-design level, and certainly not within a unify-
ing semantic paradigm. We hope our unifying foundation will facilitate the design of
such new languages.
5.3 Example: Public IVansportation
We now show an example of a program written in ambient notation. Some additional
constructs used here have been introduced in the previous section.
This example emphasizes the mobility aspects of ambients, and the fact that an am-
bient may be transported from one place to another without having to know the exact
route to be followed. A passenger on a train, for example, only needs to know the des-
tination station, and need not be concerned with the intermediate stations a train may or
may not stop at.
In this example, there are three train stations, represented by ambients: stationA,
stationB and stationC (of course, these particular ambients will never move). There are
three trains, also represented by ambients: a train from stationA to stationB originating
 88

at stationA, and two trains between stationB and stationC, one originating at each end.
There are two passengers, again represented by ambients; /oe and nancy. Joe wants to
go from StationA to stationC by changing trains at stationB; nancy wants to go the other
way.

StationA stationB . „^ stationC

joe ^ * "^
trainAB trainBC 1
We begin by defining a parametric process that can be instantiated to trains going
between different stations at different speeds. The parameters are: stationX: the origin
station; stationY: the destination station; XYatX, the tag that the train between X and Y
displays when stationed at X; XYatY, the tag that the train between X and Y displays
when stationed at Y\ tripTime, the time a train takes to travel between origin and desti-
nation.
let train{stationX stationY XYatX XYatY tripTime) =
(v moving) II assumes the train originates inside stationX
moving[rec T.
be XYatX. wait 2.0.
be moving, out stationX. wait tripTime. in stationY.
be XYatY. wait 2.0.
be moving, out stationY. wait tripTime. in stationX.
n
The definition of a train begins with the creation of a new name, moving, that is in-
termittently used as the name of the train. While the train is moving, passengers should
not be allowed to (dis)embark; this is achieved by keeping moving a secret name. The
train begins as an ambient with name moving, and contains a single recursive thread that
shuttles the train back and forth between two stations. Initially, the train declares itself
to be a train between X and Y stationed at X, and waits some time for passengers to enter
and exit. Then it becomes moving, so passengers can no longer (dis)embark. It exits the
origin station, travels for the tripTime, enters the destination station, and declares itself
to be the train between X and Yat Y. Again, passengers can (dis)embark during the wait
time at the station. Then the train becomes moving again, goes back the other way, and
then repeats the whole process.
Next we have the configuration of stations and trains. We create fresh names for
the three stations and for the train tags. Then we construct three ambients for the three
stations, each containing an appropriately instantiated train.
(v StationA stationB stationC ABatA ABatB BCatB BCatC)
stationA[train{stationA stationB ABatA ABatB 10.0)] I
stationB[train{stationB stationC BCatB BCatC 20.0)] I
stationC[train(stationC stationB BCatC BCatB 30.0)] I
 89

Finally, we have the code for the passengers. Joes itinerary is to enter stationA,
wait to enter the train from A to B when it is stationed at A, exit at B, wait for the train
from fi to C when it is stationed at B, exit at C and finally exit stationC. During the time
thatyoe is waiting to exit a train, he is blocked waiting for the train to acquire the appro-
priate tag. The train could change tags at intermediate stations, but this would not affect
joe, who is waiting to exit at a particular station. When that station is reached, and the
train assumes the right tag, joe will attempt to exit. However, there is no guarantee that
he will succeed. For example, joe may have fallen asleep, or there may be such a rush
that joe does not manage to exit the train in time. In that case,7oe keeps shuttling back
and forth between two stations until he is able to exit at the right station.
(vjoe)
joe[
in StationA.
inABatA. out ABatB.
in BCatB. out BCatC.
out StationC] I
The code for nancy is similar, except that she goes in the other direction. Given the
timing of the trains, it is very likely that nuncy will meet/oe on the platform at stationB.
(v nancy)
nancy [
in StationC.
in BCatC. out BCatB.
in ABatB. out ABatA.
out StationA]
In all this,7oe and nancy are active ambients that are being transported by other am-
bients. Sometimes they move of their own initiative, while at other times they move be-
cause their context moves. Note that there are two trains between stationB and stationC,
which assume the same names when stopped at a station. Joe and nancy do not care
which of these two train they travel on; all they need to know is the correct train tag for
their itinerary, not the "serial number" of the train that carries them. Therefore, having
multiple ambients with the same name simplifies matters.
When all these definitions are put together and activated, we obtain a real-time sim-
ulation of the system of stations, trains, and passengers. A partial trace looks like this:
nancy: moved in stationC
nancy: moved in BCatC
joe: moved in stationA
joe: moved in ABatA
joe: moved out ABatB
nancy: moved out BCatB
joe: moved in BCatB
nancy: moved in ABatB
 90

nancy: moved out/iBaf/1

nancy: moved out stationA
joe: moved out BCatC
joe: moved out stationC
5.4 Challenge: A Conference Reviewing System
We conclude with the outHne of an ambitious wide-area application. The application
described here does not fit well with simple-minded Web-based technology because of
the complex flow of active code and stateful information between different sites, and
because of an essential requirement for disconnected operation. The application fits
well within the agent paradigm, but also involves the traversal of multiple administra-
tive domains, and has security and confidentiality requirements.
This is meant both as an example of an application that could be programmed in a
wide-area language, and as a challenge for any such language to demonstrate its usabil-
ity. We hope that a language based on ambients or similar notions would cope well with
this kind of situation.
Description of the problem
The problem consists in managing a virtual program committee meeting for a confer-
ence. The basic architecture was suggested to me by comments by Richard Connors, as
well as by my own experience with organizing program committee meetings and with
using Web-based reviewing software developed for ECOOP and other conferences.
In the following scenario, the first occurrence of each of the principals involved is
shown in boldface.
Announcement
A conference is announced, and an electronic submission form, signed by the confer-
ence chair, is publicized.
Submission
Each author fetches the submission form, checks the signature of the conference chair,
and activates the form. Once activated, the form actively guides most of the reviewing
process. Each author fills an instance of the form and attaches a paper. The form checks
that none of the required fields are left blank, electronically signs the paper with a sig-
nature key provided by the author, encrypts the attached paper, and finds its way to the
program chair. The program chair collects the submissions forms, and gives them a
decryption key so that they can decrypt the attached papers and verify the signatures of
the authors. (All following communications are signed and encrypted; we omit most of
these details from now on.)
Assignment
The program chair then assigns the submissions to the committee members, by in-
structing each submission form to generate a review forms for each assigned member.
The review forms incorporate the paper (this time signed by the program chair) and find
their way to the appropriate committee members.
 91

Review
Each committee member is a reviewer, and may decide to review the paper directly, or
to send it to another reviewer. The review form keeps tracks of the chain of reviewers
so that it can find its way back when either completed or refused, and so that each re-
viewer can check the work of the subreviewers. Eventually a review is filled. The form
performs various consistency checks, such as verifying that the assigned scores are in
range and that no required fields are left blank. Then it finds its way back to the program
chair.

Report generation
Once the review forms reach the program chair, they become report forms. The vari-
ous report forms for each paper merge with each other incrementally to form a single
report form that accumulates the scores and the reviews. The program chair monitors
the report form for each paper. If the reviews are in agreement, the program chair de-
clares the form an accepted paper report form, or a rejected paper review form.

Conflict resolution
If the reports are in disagreement, the program chair declares the form an unresolved
review form. An unresolved review form circulates between the reviewers and the pro-
gram chair, accumulating further comments, until the program chair declares the paper
accepted or rejected.
Notification
The report form for an accepted or rejected paper finds its way back to the author (minus
the confidential comments), with appropriate congratulations or regrets.
Final versions
Once it reaches the author, an accepted paper report form spawns a final submission
form. In due time, the author attaches to it the final version of the paper and signs the
copyright release notice. The completed final submissions form finds its way back to
the program chair.
Proceedings
The final submission forms, upon reaching the program chair, merge themselves into
the proceedings. The program chair checks that all the final versions have arrived, sorts
them into a conference schedule, attaches a preface, and lets the proceedings find their
way to the conference chair.
Publication
The conference chair files the copyright release forms, signs the proceedings, and posts
them to public sites.
Comments
A few critical features characterize this application as particularly well-suited for exper-
imenting with wide-area languages.
First, it is a requirement of this application that most interactions happen in absence
 92

of connectivity. Virtual committee meetings occur over the span of one month of re-
viewing and one or two weeks of discussion. It is highly unlikely that all the committee
members will be continuously near their main workstation, or any workstation, during
such a span of time. Yet, progress cannot be interrupted by the temporary absence of
any one member. Furthermore, progress cannot be intenupted by the absence of con-
nectivity for any one member: paper reviews are commonly done on airplanes, in doc-
tors waiting rooms, in lines at the Post Office, in cafe's, etc. While a laptop or personal
organizer can be easily carried in those environments, continuous connectivity is far
from easy to achieve. This is to be contrasted with current web-based review systems,
which require reviewers to sit at a connected workstation while filling the review forms.
Second, form-filling requires semantic checking, which is best done while the form
is being filled. Therefore, active forms are required even during off-line operation. This
is to be contrasted with the filling of on-line Web-based review forms which require, in
practice, preparing reviews off-line on paper or in ASCII, and later typing them or past-
ing them laboriously into on-line forms in order to obtain the semantic checking. Alter-
natively, if the review is simply e-mailed in ASCII, then the program chair has the
considerable burden of performing the parsing and semantic checking.
Third, unattended operations is highly desirable also for the program chair. The
program chair may go in vacation after the assignment phase and come back to find all
the report forms already merged, thanks to the use of active forms.
Fourth, the system must handle multiple administrative domains. Committee mem-
bers are intentionally selected to belong to widely diverse and dispersed institutions,
many of which are protected by firewalls. In this respect, this situation is different from
classical office workflow on a local area network, although it shares many fundamental
features with it.
Fifth, all the forms are active. This relieves various principals from the tedious and
error-prone task of collecting, checking, and collating pieces of information, and dis-
tributing them to the correct sets of other principals.
In summary, in this example, interactions between various parts of the system hap-
pen over a wide-area network. The people involved may be physically moving during
or between interaction. As they move, they may transport without warning active parts
of the system. At other times, active parts of the system move by their own initiative
and must find a route to the appropriate principals wherever they are.

6 Conclusions
The global computational infrastructure has evolved in fundamental ways beyond stan-
dard notions of sequential, concurrent, and distributed computational models. Mobile
(imbients capture the structure and properties of wide-area networks, of mobile comput-
ing, and of mobile computation. The ambient calculus [13] formalizes these notions
simply and powerfully. It supports reasoning about mobility and security, and has an
intuitive presentation in terms of the folder calculus. On this basis, we can envision new
programming methodologies, libraries and languages for global computation.
 93

7 Acknowledgments
The ideas presented in this paper originated in the heated atmosphere of Silicon Valley
during the Web explosion, and were annealed by the cool and reflective environment of
Cambridge UK. I am deeply indebted to people in both locations, particularly to An-
drew Gordon, who is also a coauthor of related papers. In addition, Martin Abadi and
Cedric Fournet made comments and suggestions on recent drafts.

References
[ I ] Abadi, M., Foumet, C , Gonthier, G.: Secure Implementation of Channel Abstractions. Proc.
of the Thirteenth Annual IEEE Symposium on Logic in Computer Science (1998) 105-116.
[2] Abadi, M., Gordon, A.D.: A Calculus for Cryptographic Protocols: the Spi Calculus. Proc.
of the Fourth ACM Conference on Computer and Communications Security (1997) 36-47.
[3] Agha, G.A.: Actors: A Model of Concurrent Computing in Distributed Systems. MIT Press
(1986).
[4] Amadio, R.M.: An Asynchronous Model of Locality, Failure, and Process Mobility. Proc.
COORDINATION 97, Lecture Notes in Computer Science 1282, Springer Verlag (1997).
[5] Bharat, K., Cardelli, L.: Migratory Applications. Proc. of the ACM Symposium on User In-
terface Software and Technology '95 (1995) 133-142.
[6] Berry, G.: The Foundations of Esterel. In: Plotkin, G., Stirling C , Tofte, M. (eds.): Proof
Language and Interaction: Essays in Honour of Robin Milner. MIT Press (1998).
[7] Berry, G., Boudol, G.: The Chemical Abstract Machine. Theoretical Computer Science
96(0(1992)217-248.
[8] Boudol, G.: Asynchrony and the Ji-calculus. Technical Report 1702, INRIA, Sophia-Antip-
ohs(I992).
[9] Bracha, G., Toueg, S.: Asynchronous Consensus and Broadcast Protocols. I.ACM 32(4)
(1985)824-840.
[10] Brauer, W. (ed.): Net Theory and Applications. Proc. of the Advanced Course on General
Net Theory of Processes and Systems, Hamburg, 1979. Lecture Notes in Computer Science
84. Springer-Vedag (1980).
[11] Cardelli, L.: A Language with Distributed Scope. Computing Systems, 8(1), MIT Press
(1995)27-59.
[ 12] Cardelli, L., Davies, R.: Service Combinators for Web Computing. Proc. of the First Usenix
Conference on Domain Specific Languages, Santa Barbara (1997).
[13] Cardelli, L., Gordon, A.D.: Mobile Ambients. In: Nivat, M. (ed.): Foundations of Software
Science and Computational Structures, Lecture Notes in Computer Science 1378, Springer
(1998) 140-155.
[14] Cardelli, L., Gordon, A.D.: Types for Mobile Ambients. Proc. of the 26th ACM Symposium
on Principles of Programming Languages (1999) 79-92.
[15] Carnero, N., Gelernter, D.: Linda in Context. Communications of the ACM, 32(4) (1989)
444-458.
[16] Carriero, N., Gelernter, D.,Zuck,L.: Bauhaus Linda. In: Ciancarini, P., Nierstrasz.,0., Yon-
ezawa, A. (eds.): Object-Based Models and Languages for Concurrent Systems. Lecture
Notes in Computer Science 924, Springer Veriag (1995) 66-76.
 94

[17] Chandra, T.D., Toueg, S.: Unreliable Failure Detectors for Asynchronous Systems. ACM
Symposium on Principles of Distributed Computing (1991) 325-340.
[18] De Nicola, R., Ferrari, G.-L., Pugliese, R.: Locality Based Linda: Programming with Explic-
it Localities. Proc. TAPSOFT'97. Lecture Notes in Computer Science 1214, Springer Ver-
lag (1997) 712-726.
[19] Fischer, M.J., Lynch, N.A., Paterson, M.S.: Impossibility of Distributed Consensus with
One Faulty Process. J.ACM 32(2) (1985) 374-382.
[20] Fournet, C , Gonthier, G.: The Reflexive CHAM and the Join-Calculus. Proc. 23rd Annual
ACM Symposium on Principles of Programming Languages (1996) 372-385.
[21 ] Fournet, C , Gonthier, G., Levy, J.-J., Maranget, L., Remy, D.: A Calculus of Mobile Agents.
Proc. 7th International Conference on Concurrency Theory (CONCUR'96) (1996) 406-421.
[22] Fournet, C , Maranget, L.: The Join-Calculus Language - Documentation and User's Guide.
<http://pauillac.inria.fr/join/> (1997).
[23] Goshng, J., Joy, B., Steele, G.: The Java Language Specification. Addison-Wesley (1996).
[24] Hoare, C.A.R., Communicating Sequential Processes. Communications of the ACM 21(8)
(1978)666-678.
[25] Honda., K., Tokoro, M.: An Object Calculus for Asynchronous Communication. Proc.
ECOOP'91, Lecture Notes in Computer Science 521, Springer Verlag (1991) 133-147.
[26] INMOS Ltd.: occam programming manual. Prentice Hall (1984).
[27] Kistler, T., Marais, J.: WebL - A Programming Language for the Web. In Computer Net-
works and ISDN Systems, 30. Elsevier (1998) 259-270.
[28] Milner, R.: A Calculus of Communicating Systems. Lecture Notes in Computer Science 92.
Springer Verlag (1980).
[29] Milner, R.: Functions as Processes. Mathematical Structures in Computer Science 2 (1992)
119-141.
[30] Milner, R., Parrow, J., Walker, D.: A Calculus of Mobile Processes, Parts 1-2. Information
and Computation, 100(1) (1992) 1-77.
[31] Morris, J.H.: Lambda-Calculus Models of Programming Languages. Ph.D. Thesis, MIT
(Dec 1968).
[32] Palamidessi, C : Comparing the Expressive Power of the Synchronous and the Asynchro-
nous Pi-calculus. Proc. 24th ACM Symposium on Principles of Programming Languages
(1997)256-265.
[33] Sander, A., Tschudin, C.F.: Towards Mobile Cryptography. ICSI technical report 97-049,
November 1997. Proc. IEEE Symposium on Security and Privacy (1998).
[34] Sangiorgi, D.: From 7i-calculus to Higher-Order 7i-calculus - and Back. Proc. TAPSOFT
'93., Lecture Notes in Computer Science 668, Springer Verlag (1992).
[35] Stamos, J.W., Gifford, D.K.: Remote Evaluation. ACM Transactions on Programming Lan-
guages and Systems 12(4) (1990) 537-565.
[36] Stein, L.A., Lieberman, H., Ungar, D.: A Shared View of Sharing: The treaty of Orlando. In:
Kim, W., Lochowsky, F. (eds.); Object-Oriented Concepts, Applications, and Databases.
Addison-Wesley (1988) 31-48.
[37] White, J.E.: Mobile Agents. In: Bradshaw, J. (ed.): Software Agents. AAAI Press / The MIT
Press (1996).
 Type-Safe Execution of Mobile Agents
in Anonymous Networks

Matthew Hennessy and James Riely

Abstract. We study type-safety properties of open distributed systems of mobile

agents, where not all sites are known to be well-typed. We adopt the underlying
model of an anonymous network, allowing that code may be corrupted on trans-
mission and that the source of incoming code is unknowable. Nonetheless, we are
able to guarantee a weak form of type-safety at "good" sites using a mix of static
and dynamic typing.

1 Introduction
In [6] we presented a type system for controlling the use of resources in a distributed
system. The type system guarantees that resource access is always safe, in the sense that,
for example, integer channels are always used with integers and boolean channels are
always used with booleans. The type system of [6], however, requires that all agents in
the system be well-typed. In open systems, such as the internet, such global properties
are impossible to verify. In this paper, we present a type system for partially typed
networks, where only a subset of agents are assumed to be well typed.
This notion of partial typing is presented using the language DTI, from [6]. In D7t
mobile agents are modeled as threads, using a thread language based on the 7i-calculus.
Threads are located, carrying out computations at locations or sites. Located threads,
or agents interact using local channels, or resources.
In an open system, not all sites are necessarily benign. Some sites may harbor mali-
cious agents that do not respect the typing rules laid down for the use of resources. For
example, consider the system

^|(vc:chan(int)) gom.a!(/c[c])|
\mial{z[x])goz.x\{x.)\
consisting of two sites k and m. The first generates a new local channel c for transmitting
integers and makes it known to the second site m, by sending it along the channel a local
to m. Here, the value k\c] and the pattern z\x\ are dependent pairs, where the first element
represents a location and the second element represents a resource at that location. In a
benign world k could assume that any mobile agent that subsequently migrates from m
to k would only use this new channel c to transmit integers at k. However in an insecure
world m may not play according to the rules; in our example it sends an agent to k which
misuses the new resource by sending the boolean value t along it.
In this paper we formalize one strategy that sites can use to protect themselves from
such attacks. The strategy makes no assumptions about the security of the underlying
network. For example, it is not assumed that the source of a message (or agent) can be
reliably determined. We refer to such networks as anonymous networks.
 96

In the presence of anonymous networks a reasonable strategy for sites is based on

paranoia. Since the source of messages cannot be determined it is impossible to distin-
guish messages from potentially "trusted" sites; thus no site can be trusted. To protect
itself, a site must bar entry of any mobile agent that cannot be proven to use local re-
sources as intended.
The requirement that incoming agents be checked entails that the runtime semantics
is of necessity more complicated than that of [6]. Each location must retain information
about its local resources against which the incoming agents are checked. This infor-
mation is encapsulated as a location type, giving the names of the resources available
locally, together with their types; agent checking then amounts to a form of local run-
time typechecking.
The paper proceeds as follows. In the next section we review the language DTI of [6]
and recall its standard semantics. While the paper is self contained, we rely on [6] for
examples and motivation for Dn. This is followed, in Section 3, by a formal description
of the modified run-time semantics, where sites retain local information against which
incoming agents are checked.
In Section 4, we discuss the effectiveness of this runtime strategy; we show that
good sites cannot be harmed by bad sites, in the sense that local resources can not
be misused at good sites, even in systems which may contain malicious sites. This is
formalized in terms of a static typing system, for which we prove Subject Reduction
and Type Safety theorems. The paper ends with a critique of both the run-time strategy
and the typing system.

2 The Language
We present the syntax and standard semantics of DJI. For a full treatment of the lan-
guage, including many examples, see [6]. The language is a simplification and refine-
ment of that introduced in [11].

2.1 Syntax
The syntax is given in Table 1. We defer the discussion of types, T, to Section 2.3. The
syntax is parameterized with respect to the following syntactic sets, which we assume
to be disjoint:
- Var, of variables, ranged over by x-z,
- Name, of names, ranged over by a-m,
- Int, of integers, ranged over by /, and
- Bool = {t,f}, oibooleans, ranged over by bv.
We use u-w to range over the set of identifiers, Id = Var U Name. We typically use the
names a-d to refer to channels and k-m to refer to locations, although the distinction is
formally imposed by the type system. We use e to refer to names that might be of either
type. The main syntactic categories of the language are as follows:
- Threads, P-R, are (almost) terms of the ordinary polyadic Ti-calculus [8]. The thread
language includes the static combinators for composition '|' and typed restriction
 97

Table 1 Syntax
Id: u,v,w e I X
Vai. U, V / I bv I M I H'[u,,..,M„] I ((7,, .., Un)
Pat: X, Y z[Xi. (A|, .., X„)

Thread: P,Q,R ::= stop I u\{V)P \ ul{X:i:)Q \ *P

I gOM./'I if ^7= V t h e n P e l s e g
I P\Q I {y^-T)P
System: M,N ::= 0 I efP} \M\N \ {Vfe:r)N

'(ve:T) ' , a s well as constructs for movement'go^./*', output ' M ! ( V ) / ' ' , typed input
'M?(X:T) Q\ (mis)matching 'if t/ = V then P else Q' and iteration '*P'.
- Agents, £lP}, are located threads.
- Systems, M-N, are collections of agents combined using the static combinators ' | '
and '(v^c:T)'.
The output and input constructs make use of syntactic categories for values, U-
V, and patterns, X-Y, respectively. Values include variables, names, base values, and
tuples of these. A value of the form w[ui,.., u„] includes a location w and a collection
of channels Uj,.., M„ allocated at w. Patterns, X-Y, provide destructors for each type.
To be well-formed, we require that patterns be linear, i.e. each variable appear at most
once.
As an example of a system, consider the term:

llPj\{via:A){em\klR])

This system contains three agents, £ | P ] , £|(2l and klRj. The first two agents are running
at location i, the third at location k. Moreover Q and R share a private channel a of type
A, allocated at I and unknown to P.
Unlike [3, 5], agents are relatively lightweight in DTC. They are single-threaded and
can be freely split and merged using structural rules and communication. As such, they
are unnamed.

NOTATION. We adopt several notational conventions, as in [6].

- In the concrete syntax, "move" has greater binding power than composition. Thus
goi.P I Q should be read (go£.P) | Q. We adopt several standard abbreviations.
For example, we routinely drop type annotations when they are not of interest.
We omit trailing occurrences of stop and often denote tuples and other groups us-
ing a tilde. For example, we write a instead of (oi,.., a„) and {Ve:T)P instead of
(veiiT,) ..(ve„:T„)P. We also write 'if U =V then P' instead of 'if f/ = V then
P else stop' and 'if (/ 7^ V then Q' instead of '\f U = V then stop else Q.'
- We assume the standard notion of free and bound occurrences of variables and
names in systems and threads. The variables in the pattern X are bound by the input
construct u?{X) Q, the scope isQ. The name e is bound by the restrictions (v^e:T) N
and (ve:T) P, the scopes are N and P, respectively. A term with no free variables is
 98

Table 2 Standard Reduction

Structural equivalence:
(s-stop) ^stopl = 0
(s-split) iip\Q] = m\m}
(s-itr) ei*pj = eiPi\£i*p}
(s-new) ^I(ve:T)Pl = {vie:T)eiP}
(s-extr) M\(vte:T)N = (\te:T){M\N) if e ^ fn (M)
Reduction:
(r-move) eigok.Pj -^ kiPl
(r-comm) eia\{V)P} \eial(X)Qj ->
(r-eqi) eid U = U then P e\se Qj —> ilP}
(r-eq2) efif U = V then P e\se Qj —> em if(//v
/v—^yv'
(v;e:T)A'-^(v;e:T)/V'
N—^N' N = M--^M' = N'
(r-str)
M\N-^M\N' N--^N'

closed. The functions fn(A^) andfid(A') return respectively the sets of free names
and free identifiers occurring in A'.
- We also assume a standard notion substitution, where N^"/x^ denotes the capture-
avoiding substitution of u for x in N. The notation A'|7^|} generalizes this in an
obvious way as a sequence of substitutions. For example, A^II^'^j/zWU = ''*^iyzH"AI-
- In the sequel we identify terms up to renaming of bound identifiers. D

2.2 Standard Reduction

The standard reduction semantics is given in Table 2. The structural equivalence {N =
M) and reduction relation (N —>• N') both relate closed system terms. The structural
equivalence is defined to be the least equivalence relation that is closed under composi-
tion and restriction, satisfies the monoid laws for composition,' and in addition satisfies
the axioms given in Table 2.
The structural rules allow for agent splitting £[P | 21 = ^{Pj I ^I2l and garbage
collection .^[stop| = 0. Note that when a new name is lifted out of an agent, using the
structural rule (s-new), the system-level restriction records the name of the location
which allocated the name £[(ve:T)/'] = (v^e:T)£[[/']; these location tags are used only
for static typing.
Most of the rules of the reduction relation are taken directly from the 7i-calculus,
with a few changes to accommodate the fact that agents are explicitly located. The
main rule of interest here is (r-move),

eigok.pj^kiPj
The monoid laws are: W10 =/V, iV IM = A/! W, and A'I (WIO) = (/VIM) I O.
 99

which states that an agent located at £ can move to k using the move operator gok.P.
This is the only rule that varies significantly between the standard semantics and the
semantics for open systems defined later. Note that in (r-comm), communication is
purely local:
4a!(v)p] I eiaiix)Qj^m \ mvm
As an example, suppose that we wish to write a system of two agents, one at k and
one at I. The agent at k wishes to send a fresh integer channel a, located at k, to the
other agent using the channel b, located at £. This system could be written:

mnz[x])Ql \kl{va){P\ goth\{k[am

= llbl{z[x]) GI I (v,a) {klPj I klgoi.h\{k[a])}) (s-new), (s-split)
-^ eibl{z[x])Q} I (v,fl) {klPj I iMk[am) (r-move)
- ^ (v*fl) (£[Gf 1"1AWH I klPj) (s-extr), (r-comm), (s-stop)
Beside each reduction, we have written the rules used to infer it, omitting (r-str) and
(r-new), which are almost always used. Note that after arriving at £, the agent sends
the value k[a] rather than simply a. In the type system, this identifies the resource a as
non-local at (.. If a simple resource value, such as a, had been communicated to Q, it
would have had to have been local to £, rather than k. An example of a process Q that
uses the received value z[x\ is goz.x!(l), which after the communication would become
gok.a\{l). This can move to the location k and send the integer 1 on the newly received
channel a.

2.3 Types and Subtyping

The purpose of the type system is to ensure proper use of base types, channels and
locations. In this paper we use the simple type languages from [6, §4], extended with
base types for integers and booleans. We use uppercase Roman letters to range over
types, whose syntax is as follows:

TChan A,B,C ::= chan(T)

TLoc K,L,M ::= loc{ai:Ai, ..,fl„:A„}, a, distinct
TVal S,T ::= bool | int | A | K[B,,..,Bft], | (T,,..,T„)

Types are divided into the following syntactic groups:

- TChan of channel types, which specify the type of values communicated over a
channel, chan(T).
- TLoc of simple location types, which specify the set of typed channels available at
a location, loc{a:A}.
- TVal of value types, which include types for base values, channels, locations and
tuples.
In value types, location types have the form loc{ai:A|,.., fl„:A„}[Bi,.., B^]. The ex-
tended form allows for a certain amount of first-order existential polymorphism. Infor-
mally, loc{2:A}[B] mayjbe read "3x: loc{2:A,jf:B}", i.e. the type of a location which
has channels a of types A and some (unnamed) channels of types B.
 100

Location types are essentially the same as standard record types, and we identify
location types up to reordering of their "fields". Thus loc{a:A,fe:B}[C]= \oc{b:B,
a:A}[C]. But reordering is not allowed on "abstract" fields. Thus if B and C are dif-
ferent, then loc{fl:A}[B,C] yi loc{a:A}[C,B].
Throughout the text, we drop empty braces when clear from context, writing 'loc'
instead of 'loc{ }[]', 'K' instead of 'K[]', and '«' instead of'«[]'.
The subtyping relation (T < S) is discussed at length in [6]. On base types and
channel types there is no nontrivial subtyping; for example, chan(T) < chan(T') if and
only if T = T'. On location types (both simple and "existential"), the subtyping relation
is similar to that traditionally defined for record or object types:

\oc{a:A, b:B] < loc{2:A}

\oc{a:A, b:B}[C] <\oc{a:A}[C\

On tuples, the definition is by homomorphic extension: S < T if V/: S, < T,-

Both subtyping and existential dependent types are inherited from our earlier work.
Whereas subtyping is not essential to the work at hand, dependent types are, as will be
seen in Section 3.1,

2.4 TVpe Environments

Location types contain the names of the channels known to be defined at a location.
To present typing systems for the language in later sections, it is useful to generalize
location types to allow the inclusion of variables as well as names. Variables are allowed
at types int and bool, in addition to channel types A. The resulting types are called v-
open location types, K:

TSimple: H,G ::= int | bool | A

TOpen: K,'L,M ::= IOC{M:H}, M,-distinct

To be well-formed, we require that every name in a v-open location type be associated

with a channel type.
The subtyping relation extends directly to v-open location types:

IOC{M:H, v:G} < IOC{M:H}

A type environment, F, maps identifiers to v-open location types. An example of a

type environment is:

{
k : loc {a:chan(int),jc:int}
^ .|ocf<^:chan(loc[chan(int)]) "i
\3':chan{loc[chan(bool)]) /

Here we have two locations, k and z- The first has an integer channel named a and an
integer variable x. The second has two channels: a, which communicates (potentially
remote) integer channels and y which communicates (potentially remote) boolean chan-
nels.
 101

If a type environment contains no variables, we say that it is closed. Closed type

environments map names to (closed) location types K.
In the typing system of the next section we need some notation for extending (v-
open) location types and for convenience we explain this notation here. The location
type L may be extended using the notation 'L, V:T', defined by induction on V:

(IOC{M:H}), v:G = loc{«;H, w:G} ifVi: M,: ^ v

L, VV[M]:K[A] = L if {w,«} disjoint dom(L)
L, bvrbool = L
L, /:int = L
L, t/:T = ((... (L,(/,:T,),...),f/„:T„)

Therefore the extension of L by V-.T adds only information about local identifiers to L.
For example:

loc{rf:D},(0,A:,z[>']):(int,A,loc{c:C}[B]) = loc{<i:D,x:A}

Note that every location type L is also a v-open location type, and thus this definition
applies to "closed" location types as well. In the same way the definition applies to
patterns, as well as values, since syntactically every pattern is also a value.
We use a similar notation for extending type environments: If M is not in the domain
of r then r, M:L denotes the new type environment, which is similar to T but in addition
maps u to type L.

3 Semantics for Open Systems

As we have explained in the introduction, using the standard run-time semanfics local
resources may be intentionally misused by malicious sites; in the example given in the
Introduction, we have seen that the local resource a at location k is deliberately misused
by a mobile agent originating at location m. In this section we modify the run-time
semantics in a manner which offers protection against such misuse.
We are assuming that agents are working in an anonymous network and the run-time
strategy suggested in the introduction is for every site to keep a record of local resources
against which all incoming agents are checked. In our formulation this distiibuted in-
formafion is collected together as closed type environment and thus reductions of the
runtime semantics take the form

where A is a closed type environment and A' and A'' are systems.
The reduction semantics is given in Table 3. Most of the rules are simple adaptations
of the corresponding rules in Table 2. For example, the rules for local communication
and matching of values are essentially as before, as A is not consulted for these reduc-
tions. There is a minor change in the rule for the restriction operator, because A must be
augmented to reflect the addition of the new name.
 102

Table 3 Reduction for Open Systems

(r-move) A^eigok.Pj ^ klPj \i'A{k) \-P
(r-comm) A>eia\(V)Pj | £[a?(X)ei ^ IfP} \ mV/xM
(r-eqi) Ao^if (7 = t/then Pelse g] -^ £IPJ
(r-eq2) A > ^Jif (/ = V then P else g] —> ^[21 if (/ / V
A,£:(L,a:A)i>A'—>A'' A, £:L, A:;K> N —> M
A, £:Lt> (vta:A)N -^ {via:A)N' A,e:L> {vik:K)N —> {vtk:K)N'
A>W^/V' WHM A>M—^W M' = N'
(r-str)
A>M\N—^M\N' A>N-^N'

The only significant change from the standard run-time semantics is in the rule for
code movement:
A > ilgok. Pj h-^- klPj if A(^) h P
This says that the agent P can move from location i to location k only if P is guaranteed
not to misuse the local resources of ^, i.e. A{k) h f. Here P is type-checked dynamically
against A{k), which gives the names and types of the resources available at k.

3.1 Runtime IVping

The definition of this runtime local type-checking is given in Table 4. This is a light
weight typing in that the incoming code is only checked to the extent of its references
to local resources. Thus judgments are of the forni

L \-P

indicating that P can safely run at a location that provides resources as defined in L.
Perhaps the most surprising rule in this light weight type checking is (t-move),
which involves no type checking whatsoever. However this is reasonable as an agent
such as go£.P running at k uses no local resources; it moves immediately to the site £.
As a result of this rule notice that reductions of the form

A> mlgok.goi.Pj H-> klgoe.Pj

are always allowed, regardless of the information in A.

The only significant local checking is carried out by the two rules (t-r), (t-w) which
we examine in some detail. The subtlety in the read rule (t-r) is to some extent hidden
in the rule for updating location types and this is best explained by example. The rule
dictates, for example, that the agent a'?{x.char\{T))Q can migrate to a location with
local resources L provided:
- a is a local channel of the appropriate type, in this case a channel for communicating
values of type chan(T)
- Qis locally well-typed with respect to an augmented set of resources, L,j::chan{T).
 103

Table 4 Typing for Values and Threads

Values:
L<IOC{M:H}
(t-sit) — = ^ (t-base)
LI-M:H L I- n:int, bv:bool
V/: L h Ui-.li
(t-loc) (t-tup)
L l-«[vi,..,v„]:K[A,,..,A„] Lh(f/,,..,f;„):(T„..,T„)
Threads:

Ll-P
t-movej (t-newl)
L \-gou.P LI-(v/t:K)/'
L h «:chan(T) L, A':T h Q L, a:A h P
(t-r) t-newc
LI-M?(X:T)2 LI-(va:A)P
LI-M:chan(T>, V : T , / ' LhP.G
(t-w) t-str
h\-u\{y)P L h stop, *P, PIG
L 1- /7:T, V:T, P, Q
(t-eq)
Lf-if ( / = K then P else g

However suppose the channel a communicates non-local information; e.g. when is

fl?(z[x]:K[A]) Q locally well-typed? The rule (t-r) simply demands that, in addition to a
having the appropriate local type, Q is well-typed with respect to the same set of local
resources L. Formally this is because according to the definitions given in Section 2.4
L,z[x]:K[A] is simply L. Intuitively this is reasonable since any non-local information
received on a will not be used locally and thus it may be ignored.
The rule (t-w) states that agent a\{V) P is locally well-typed provided

- the continuation P is locally well-typed

- the channel a has an appropriate local type, say chan(T)
- the value to be transmitted V is locally well-typed to be transmitted on a, L I- V:T;
this means that it must be possible to assign to V the local object type of the channel
a, namely T.

Once more there is a subtlety, this time in the local type checking of values. If the value
V to be transmitted is a local resource, say a channel name b, then according to the rule
(t-sit) b must have the local type T. If, on the other hand, V is a non-local value, say
k[b], then locally this is of no interest; according to (t-loc) k.[b] can be assigned any
location type which in effect means that when it is transmitted locally on a its validity
is not checked.
This ends our discussion of runtime local type checking, and of the runtime seman-
tics.
 104

3.2 An Example
As an example consider a system of three locations, k, i and m, with the following
distributed type environment, A.

{
k : loc{a:chan(int)}
i : loc{i>:chan(loc[chan(bool)])}
m : loc{(3f:chan(loc[chan{bool)])}

Let A^ be the following system;

klgom.d\{k[a])j
\mldnz[x]) goLb\{z[x])j
I ilb?iz[x])goz.x\{t)j
Here k communicates the name of its integer channel a to m, using the channel d local
to m. Then m misinforms i about the type of a at t. the communication along b fools (.
into believing that a is a boolean channel. Subsequently i attempts to send an agent to
k that violates the type of local resource a, by sending a boolean value where an integer
is expected.
The reader can check that according to our runtime semantics the first code move-
ment between k and m is allowed:
A> klgom.d]{k[a])}^^ mld\{k[a\)j

as local type checking of the migrating agent succeeds, A(m) h d\{k[a]).The local chan-
nel d is used correctly and since the value transmitted, k[a], is non-local it is essentially
not examined (only the number of names is checked, not their types).
The local communication at m on channel d now occurs and the second code move-
ment between m and £ is also allowed,

At>mld\{k[a])j I mldl{z[x]) go£.b\{z[xm^ •^ eib\{k[a])j

because the migrating thread, b\{k[a\), is also successful in its type check against local
resources, A{£). The local communication along b now occurs

A> ilb\{k[a])j I ilh-?{z[x])goz.xl{t)}^£lgok.a\{t)j

However the next potential move, the migration of the thread a! (t) from £ to /:, is disal-
lowed by the rule (r-move); the thread is locally type checked against the resources at
k, where a is known to be an integer channel, and its potential misuse is discovered.

4 Static typing
In the runtime semantics misuse of local resources can certainly occur, since there is no
requirement that values have the object type specified by the transmitting channel. For
example the reduction

A > ^|a!(t>l Iklalix) Qj ^ ^Istop] | kMUl

 105

Table 5 Static Typing

Values and threads: As in Table 4
Systems:
hhP £ i dom(r)
(t-rung) (t-runb)
r, c-.h h £IP} r^m
r,E:(L,a:A) hN e ^ doni(r) r\-N
(t-newcg) (t-newcb)
rj:L\-{V(a:A)N rh{\/(a:A)N
r, e-.L, t.K \-N i ^ dom(r) rh-N
(t-newig) (t-newlb)
r,t.h}-{vek:K)N rh(\ik:K)N
r\-M,N
(t-str) -
r\-0,M\N
Configurations:
rh/v
(t-config) V^ e dom(r): r(£) < A{k)
r\-At>N

is allowed even if the object type of the channel a at ^ is int, i.e. A{k) < loc{a:chan(int)}.
We do not assume that all sites respect the typing constraints on their channels. For
convenience let us call sites that violate the typing constraints bad sites, as they do not
play according to the rules. A typical example is the site m, described at the end of
the previous section; it receives an integer channel a from k but then attempts to use
a to send a boolean value. In contrast a good site is one where typing constraints are
enforced.
In this section we present a static type system that guarantees that:
good sites cannot be harmed by had sites.
That is, local resources at a good site cannot be misused despite the existence of, and in-
teraction with, bad sites. We prove Subject Reduction and Type Safety theorems for the
type system. Intuitively, Subject Reduction can be interpreted as saying that the integrity
of good sites is maintained as computation proceeds, while Type Safety demonstrates
that local resources at good sites cannot be misused.
The static typing relation for anonymous networks is defined in Table 5. Judgments
are of the form
ri-A'
where F is a (v-open) type environment and A' a system. The type environment only
records the types of good locations; thus k £ dom(r) is to be read "k is good" and
m ^ dom(r) may be read as "m is bad." If F h N, then those agents in A' that are located
at sites in the domain of F are guaranteed to be "well behaved".
For threads and values, the static typing relation is the same as runtime typing rela-
tion given in Table 4. The typing of a located agent £1^1 depends on whether £ is good.
 106

i.e. i 6 dom(r). If I is good, then according to (t-rung), to infer F h t\P\ we need

to establish that the P is well-typed to run at i, r(£) 1- P. But if i is not good, then it
doesn'tmatter whetherP is well-typed; thus (t-runb) says that if £ ^ doin(r), then i\P\
is well-typed for any thread P.
The typing of new resources at the systems level also depends on whether they
are introduced at a good site. For example to infer F, £:L h (V(a:A)N, where because
of the notation we know £ is a good site, (t-newcg) says that we need to establish
F, i:{h, a:A) h N, i.e. that A' is well-typed under the assumption that there is a new
resource a at site i. This is in contrast to the case when t. is not good, £ ^ dom(F),
where according to (t-newcb) it is sufficient to establish that the system A^ is well-typed
relative to the same typing environment. Note that implicit in these two rules is the
assumption that new good sites can only be generated by good sites.
Intuitively T \- N means that in A' the good sites, those in the domain of F use
their local resources in accordance with F, whereas the behavior of bad sites are uncon-
strained. As an example consider the configuration A > A' from Section 3.2. If we let F
be the typing environment defined by:

„ _ J ^ : loc{a:chan(int)} 1 ,^,
\ £ : loc{Z7:chan(loc[chan(bool)])} J ^'

Then one can easily check that F h A > A', that is A is consistent with F and F h A^.
Intuitively here we are saying that k and £ are good sites which use their local resources
correctly whereas no guarantee is made about the local behavior at m. Note, however,
that if one considers static typing under A, which includes m, then A does not type A'.
The static typing system satisfies several important properties, such as type special-
ization and narrowing, which are stated and proved in Appendix A. These properties
are used to establish the following result, which states that well-typing at good sites is
preserved by reduction.
THEOREM 1 (SUBJECT REDUCTION).
IfT Y- A> N and A> N ^^ N' thenV \- A\> N'.
Proof. See Appendix A. D

We now discuss the extent to which in the type system precludes the misuse of lo-
cal resources at good sites. This can be formalized using a notion of runtime error. In
this paper, we confine our attention to runtime errors based on arity mismatching. Intu-
itively, an arity mismatch occurs when the value sent on a channel does not match the
type that the recipient expects, or when two structurally dissimilar values are compared
using the match construct. To formalize this notion, we define a compatibility relation
(/ X T between (closed) values and types, and a compatibility relation [/ x V between
(closed) values.
The definitions of compatibility and of runtime error are given in Table 6. Runtime
error is defined as a predicate A' - ^ ^ , indicating that A' is capable of a runtime error at
location £. Essentially an error occurs at location £ if either two incompatible values are
compared at £ or an attempt is made to communicate a value along a local channel which
is incompatible with the type of the channel. The only non-trivial rule is (e-new) which
 107

Table 6 Compatibility and Runtime Error

Compatibility:
i X int i X i'
bv X bool bv X bv'
e XA exe'
e[d,, ..,d„] X K[A|, .., A„] t'[^,,„,^„]xe'K,..,<]
(V,,.., V „ ) x ( T , , . . , T „ ) (K,,..,v„)x(K,',..,v;,)
ifV(: V , x T , ifVj: V,x V;
Runtime error:
(e-comm) ela\{V)Pj \ £la'!iX:T)Q} - ^ ifV^T
(e-eq) £|if ( / = K then P else g ] - ^ if f/ ^ V
N -^^ ^.m:^ yy^^ M - ^
(C new; \c • " • ' M I N - ^ ^ N - ^

may effect a change in the name of the location where the error occurs. For example it
will report an error at £ in {vit.K) N if there is a runtime error in A' at location k.
We should point out that this definition of run-time error is considerably weaker
than that employed in [6]; in that paper, the notion of run-time error took into account
not only to arity mismatches but also access violations.
THEOREM 2 (TYPE SAFETY). ifT \-N andi£ dom(r) then N ^^^.
Proof. See Appendix A. D
This theorem, together with Subject Reduction, can be interpreted informally as saying
that as reductions proceed local resources cannot be locally misused at good sites, even
in systems where not all sites which are necessarily well-behaved.

5 Discussion
Here we make three points which demonstrate the limitations of both the runtime strat-
egy and the typing system.
First the static typing is very weak as it is designed only to eliminate local misuse
of local resources. Let

{
„ I r/7:chan(loc{J:chan(bool)}) "1
I c:chan(loc{(f:chan(int)}) J
/::loc{<i:chan(int}}

Then F h £[c?(z[x])Z7!(2[jc])J although obviously here there is, at least informally, a

misuse of local channels. However formally the values exchanged on c and b are poten-
tially remote and therefore these values are not considered local. For example, consider
a companion site k which has one local channel d of type chan(int). Then no runtime
error occurs in the system
/£Ic?(z:loc{c(:chan(int)})/j!(z)l
''\\klgoL{c\{k)bl{z)goz.d\{xm
 108

although one could argue that at location I there is a misuse of the channel b. A runtime
error is avoided by the dynamic type-checking; after the communications on c and b the
potential move
ilgok.d\{t)]^kld\{t)j
is blocked because the agent attempting to move to k, d\{t) does not type check against
the local resources at k.
Second, the requirement to dynamically type-check all incoming threads is very
inefficient; however, the weak typing system makes it is essential. Purely local type-
checking makes it very difficult to introduce trust into the system. A site cannot even
trust itself! For example suppose we revised the reduction rule (r-move) to read as fol-
lows:
A > llgok. P] h-> A > klPj if k = E or A{k) h P
Here the site k trusts itself and therefore does not type check the thread P. However
this rule is not safe; Subject Reduction fails and runtime errors may be introduced. As
an example, consider the following configurafion, which uses the typing environment F
given at the beginning of this section:

F^4c?(z)goz.d(t)l \eic\m

The configuration can be typed with respect to F itself, but after the communication
the result is F o £|c!(t)|, which fails to type under F. Moreover F > £|c!(t)] induces
a runtime error due to the potential misuse of channel c. A related phenomenon is the
potential misuse of channel names as locations; e.g. the thread '(va:chan{int))goa./''
is typable in our system.
Third, the system relies heavily on dynamic type-checking to avoid misuse of re-
sources. This can be highlighted by considering another desirable property of a runtime
semantics for open systems:
Movement between good sites should always be allowed, even in the presence
of badly behaved sites.
It is obvious from the example discussed in Section 3.2 that our run-time semantics does
not satisfy this property. If we use the static environment F defined in (*) (Section 4)
then k and £ are good sites. But as we have seen in Section 3.2 the intervention of m
eventually prevents a movement from k and £.
In a companion paper [12] we address these issues by strengthening the typing sys-
tem. The second concern raised here (the inability to trust oneself or to ensure that
channels are not used as locations) can be addressed by reformalizing the typing sys-
tem while retaining its basic "local" character; such a reformalization is given in the
appendix of [12].
In [12], we go further, however, extending the notion of trust to arbitrary collec-
tions of sites. Dynamic typechecking is strengthened by making a site record informa-
tion about all sites in the system, not only itself. While typechecking in this system is
computationally more expensive, not all incoming threads need be checked; those orig-
inating at trusted sites are allowed through unchecked. This new semantics addresses
the first two of the above concerns fully, and the third partially: a stronger notion of
 109

type safety is guaranteed, some incoming threads are not typecheciced, and movement
between mutually trusted sites is always allowed, although movement between good
sites may not be.

6 Related Work

In this paper we have outlined a strategy for ensuring that the integrity of well-behaved
sites is not compromised by the presence of potentially malicious mobile agents. More-
over we have formalized the correctness of this strategy in terms of Subject Reduction
and Type Safety theorems for a partial type system.
In this study we used the language Dn [6], one of a number of distributed versions of
the 7t-calculus [8]. For other variations see [5, 13]. The languages in [3, 4] are themati-
cally similar although based on somewhat different principles. We have taken advantage
of a rich type system for D7t, originally presented in [6], where not only do channels
have the types originally proposed in [10] for the Ji-calculus, but locations have types
broadly similar to those of objects. An even richer type system is also proposed in [6]
in which types correspond to capabilities, as in [4], and an interesting topic for future
research would be the extension of partial typing to these richer types.
Our research is related to proposals for proof-carrying code outlined in [9]: code
consumers, which in our case are locations, demand of code producers, in our case
incoming threads, that their code is accompanied by a proof of correctness. This proof
is checked by the consumer before the code is allowed to execute. The correctness
is expressed in terms of a public safety policy announced by the consumer and the
producer must provide code along with a proof that it satisfies this policy. In our case
this safety policy is determined by the location type which records the types of the
consumer's resources, and proof checking corresponds to type checking the incoming
code against this record. Our work is different in that the correctness proof can be
reconstructed efficiently, and therefore the producer need not supply an explicit proof.
For other examples of related work within this framework see [7, 14]. For example
the former contains a number of schemes for typechecking incoming code for access
violations to local private resources. However the language is very different from ours,
namely a sequential higher-order functional language, and there is no direct formaliza-
don of the fact that distributed systems which employ these schemes are well-behaved.
A very different approach to system security is based on the use of cryptography
and signatures. For example [1] presents a 7i-calculus based language which contain
cryptographic constructs which ensure the exchange of data between trusted agents,
while [2] contains a description of the application of this approach in a practical setting.

A Proofs

A.l Properties of the static type system

First we prove two important properties of type systems with subtyping; Type Special-
ization and Weakening.
PROPOSITION 3 (TYPE SPECIALIZATION). Ifh \- V-.Tand! <S then L I- V:S.
 110

Proof. By induction on the judgment L h V:T. If V:T takes the form VrH then S must
coincide with H, since there is no non-trivial subtyping on channel types or base types.
If V:T has the form H'[M]:L[A] then the result is trivial, using (t-loc). Finally, the case
for tuples follows by induction. D

PROPOSITION 4 (WEAKENING).
- / / L h V:T andK<l. then IK h V:T
- / / L h P and K < L then K\- P
- Ifr, w.L \-NandK<L then r,w:K\-N
Proof. In each case the proof is by induction on the type inference. We examine two
examples of proof on threads:
(t-r). Here L h «?(X:T)Q because L I- u;chan(T) and L, X:T I- Q. We can apply the
first statement in the proposition to the former, to obtain K h M:chan(T), while
induction to the latter gives K, X:J h Q. An application of (t-r) now gives the
required K h M?(X:T) Q.
(t-newc). Here L I- {va:A)P because L, a.A \- P. By a-conversion we can choose a
so that it does not appear in K and therefore by induction we have K, a.A h P. Now
an application of (t-newc) gives the required K I- {va:A)P.
We present four cases for the proof on systems.
(t-rung). Here r,w:L (- m[P] because M h P, where M = (r,H':L)(TO). If m and w
are different then we also have M = (r,H':K)(m) and therefore an application of
(t-rung) gives the required FiVvrK h m|P|. On the other hand if m is the same as
w then M = L. So we can apply the second part of the proposition to M, obtaining
K h P. Now (t-rung) also gives the required F, w.K h m|P].
(t-runb). This case is trivial.
(t-newig). Here T, w:L h (v^m:M) W because £ e dom(r, w.h) and F, w:L, m:M H A^.
Applying induction we obtain F, w.K, m:M h A'. Now (t-newIg) can be applied
since i £ dom(F, w.K), to obtain the required F, w.K h {V(m:M)N.
(t-newcb). Here r,w.h h {V(a:A)N because £ ^ dom(F,w;L) and F,w:L h A^. How-
ever we also have £ ^ dom(F, w.K) and therefore (t-newcb) can also be applied to
obtain the required FjWiK h (V(a:A)N. D
The following Restriction Lemma states that if F h A' and some identifier u does
not occur free in A' then N can also be typed in an environment obtained from F by
removing all occurrences of «. For any identifier u let F\M denote the result of removing
all occurrences of u from F. For example (F, u:L) \M denotes F while (F, w. (L, u:A))\u
is the same as (F\«), w.h.
LEMMA 5 (RESTRICTION).
- If h,v:H h U:Tandv^ t\d{U) then L h t/:T
- If h,v:H \-Pand V ^ M(P) then L h P
- Ifr \-N and v^M(N) then r\v\-N.
Proof. By induction on the proof of the typing judgment. D
The following corollary follows by an easy induction on V.
 111

COROLLARY 6.

- //L,V:S h U:Tandf\d{V) disjoint fid (t/) thenL h U:T

- //L, V:S \- P andfid(V) disjoint M{P) then L h P

A.2 Subject Reduction

We first show that typing is preserved by the structural equivalence. The most compli-
cated case is already covered by the Restriction proposition.

LEMMA 7 (SCOPE EXTRUSION).

Ife i fn(M) then T\-M\ (v^e:T) if and only ifT h (v^e:T) (M | A')

Proof. There are two cases. If £ ^ dom(r), then we can reason as follows:

r h (v^e:T) (M I/V) 4=» T \-N mAT h M

rf-Af|(v^e:T)A/

In the case that (. € dom(r), the argument is slightly different depending on whether
e is a channel or a location. As an example we consider the former, and we assume T
has the form A,£;L.

r K {via:K){M\N) <=^ A, £:(L, a:A) h{M\N)

A, t.{h, a:A) h M and A, €:(L, a:A) h A'
A, e-.L h M and A, £:(L, a:A) \-N by Restriction
r(-Mandn-(v^a:A)7V
r[-M\{vea:A)N D

PROPOSITION 8. IfN = M then T \- N implies T \- M

Proof By induction on the proof that N = M. All of the rules and most of the axioms
are very easy to handle; the most difficult case is (s-extr), which follows by the previous
lemma. As an example, we consider the rule (s-new): £|(ve:T)P] = (v^e:T)£|/'].
If £ ^ dom(r) then it is immediate as both sides trivially type with respect to F. So
suppose £ e dom(r), and, as an example, that e:T is a:A. Let T have the form A,£:L.
Then T h ei{va:A)Pj if and only if L,fl:A h P while

ri-(v;a:A)£lPl <^=^ A,£;(L,a:A) h £[Pl

<;=> L,a:A \-P 0

As is normally the case the proof of Subject Reduction depends on the fact that, in some
sense, typing is preserved by substitution. To prove this fact the following lemma will
be useful:

LEMMA 9. / / L I- V:S andh,X:S h U:T then L h i7|7A-[}:T.

 112

Proof. By induction on the structure of X.

Suppose X is a variable, i.e. X = x for some x. We then proceed by induction on U.
If [/ is a base value or has the form M[V], then the result is trivial. If (/ is a tuple we
can apply the inner induction hypothesis. If U is x, then ^17^1} = ^ 'ind, by the typing
rules, S < T; using type specialization, we can therefore conclude L h [J^^/x^:T, as
required. If U is an identifier u that is different from x then x ^ fid(i/) and i/flYfJ = U;
therefore by the Restriction Lemma L h L'{] YA|}:T, as required.
Suppose X has the form z\x\, so L,X:S = L and fid(X) disjoint dom(L) and therefore
[/{|7x| = U. By Corollary 6, L h U:T.
If X is a tuple, the result follows by induction. D

PROPOSITION 10 (SUBSTITUTION).//L h V:Tandh,X:T \-Pthenh \-P\^/xl.

Proof. For convenience we use P' to denote P{] ^/x^ (and employ similar notation for
other syntactic categories).
As before, the proof is by induction on the structure of X. The cases for tuples Y and
structured values z\x\ are as before. We present the case where X = x, for some variable
x. In this case we proceed by a second induction on the inference L,x:T h P. We present
the cases that involve binders.
(t-r). Here L,A::T \- M?(K:S)(2 because L,x:T \- M;chan{T) and L,JC:T, Y:S \- Q. Ap-
plying the previous lemma to L,x:T h M:chan(T), we obtain L h M':chan(T). In
addition, L,x:T, Y:S h Q can be rewritten as L, K:S,J;:T h Q. By the internal in-
duction, we obtain L, Y:S h Q'. Now (t-r) gives the required L h u'l{Y:S) Q'.
(t-newc). Here L,x:T \- {va:A)P because L,x:T, a:A h P. The environment can be
rewritten as L, a.A,x:T, and therefore induction can be applied to obtain L, a: H P'.
The result follows by an application of (t-newc). D

THEOREM (1, SUBJECT REDUCTION).

IfT[-A>NandAt>N^~^ N' then Th Av>N'.
Proof From the hypothesis F h A > A' we know that F h A' and T{t) < A{£) for every £
indom(r). So it is sufficient to show F h A'', which we do by induction on the derivation
of A > A' I—> N'. We present a number of representative cases.
(r-move). Here we have A c llgok.P} —)• klP} and A{k) \- P.lfk^ dom(F) then the
required F h klPj follows trivially from (t-runb). If ^ € dom(F) then we know that
r{k) < A{k) and therefore, by Weakening, r{k) \- P. Now (t-rung) can be applied
to obtain rf-A:[/'I.
(r-comm). Here we have

A>ila\{V)Pj I ^[a?(X:T)e] - > £ M | 4 2 ^ / 4 1

If i ^ dom(F) then the result is trivial from (t-runb). Otherwise £ 6 dom(F). From
F h ila\{V)Pj I ila7{X) Qj we know F h iia\{V)Pj and therefore F(£) h P. It
follows that F h ^ l P j .
It remains to show that F h ^IgflV^lli that is F(^) h !2{|7x[}. Again from the
hypothesis we know F h £|a?(X:T) Qj from which we can conclude that F(£) h
M:chan(T) and L, X:T h Q. From F h £[[«!{V)/'J we know that r(£) h V:S for
 113

some S for which we also have r(£) h M:chan(S). In our typing system this must
mean that S and T coincide. We may therefore apply the Substitution lemma to
obtain the required r(^) h Q{|7^|}-
(r-new). We consider the case:
^,^•.\^>{vta•.A)N-^{V(a:A)N' because A, £:(L, a:A) t> A^-^^ A''
First suppose £ € dom(r). Since F h A, £:L o {vea:A)N we know F can be written
as F', i:L', where L' < L and therefore F', £:(L', a:A) h N. We can now apply in-
duction to obtain V, £:(L', a:A) h A'', to which (t-newcg) can be applied to obtain
the required F h {V(a:A)N'.
If £ ^ dom(F) then by (t-newcb) it is sufficient to prove F h A^'. In this case F I-
A, £:L > {\i(a:A)N yields F h A, £:(L, a:A) > N to which induction can be applied
to give the required T \- N'.
(r-str). This case follows using induction and Proposition 8. D

A.3 Type Safety

We first show that the typing system is "compatible" with the compatibility relation x .
LEMMA 11.
- L h V:T implies V x T
- L h V:T and L h L':T implies V x {/
Proof. A straightforward inductive argument, in the first case on the derivation of L h
V:T and in the second on the structure of the type T. D
THEOREM (2, TYPE SAFETY). / / F \-N andiG dom(F) then N H ^ .
Proof. By induction on the proof that N i-£^, we show that if £ G dom(r) and N H ^ ^
then F F A', which is sufficient to establish the theorem. Let L denote F(£).
(e-comm). In this case we have £[a!(V) Pi | £|a?(A':T) Qj H ^ because V ^ T. Now
suppose, for a contradiction, that F h £|fl!(K>P| | £lal{X:T) Qj. Because there
is no subtyping on channel types we know L (- chan(T) and L h chan(S) implies
T = S. From the alleged typing we can therefore conclude that L h V:T. By the first
part of Lemma 11, we have V x T which contradictions V :J^T.
(e-eq). Here we have £|if U =V then P else QJ H ^ ^ because U ^ V. If we assume
F h £Iif U=V then P else Qj then we must have F h U:T and F h V:T for some
T. Now applying the second part of Lemma 11 we obtain a contradiction to U :^V.
(e-new). First suppose that (VK:E) A' H ^ because A^ t - ^ and e^Llfk^ dom(F)
then by induction we have F ^ A' and by either (t-newcb) or (t-newlb) we can
conclude F ) / {Vke-B)N.
So suppose k G dom(F). As an example suppose e:E is a location; the case when
it is a channel is similar. Applying induction we have F,e:E Y N and therefore by
(t-newig) we can conclude F Y (v^-eiE) A/.
Finally suppose {Vfk:E)N i-^^^ because N i '"'''^>. By a-conversion we can as-
sume that k does not appear in F and so we can apply induction to obtain r,^:E /
h A'. We know that £ € dom(r) and therefore by (t-newIg) we can conclude F Y
{vek:E)N.
(e-str). Straightforward, using induction and Proposition 8. D
 114

Acknowledgements

We thank the referees for several comments that sharpened the presentation. Matthew
Hennessy funded by CONFER II and EPSRC project GR/K60701. James Riely funded
by NSF grant EIA-9805604.

References

[1] M. Abadi and A.D. Gordon. A calculus for cryptographic protocols: The spi calculus.
Information and Computation, To appear. Available as Compaq SRC Research Report 149
(1998).
[2] B. Bershad, B. Savage, R Pardyak, E. Sirer, D. Becker, M. Fiuczynski, C. Chambers, and
S. Eggers. Extensibility, safety and performance in the SPIN operating system. In Sympo-
sium on Operating Systems Principles, pages 267-284, 1997.
[3] L. Cardelli and A.D. Gordon. Mobile ambients. In Foundations of Software Science and
Computational Structures, volume 1378 of Lecture Notes in Computer Science, pages 140-
155. BerHn: Springer-Verlag, 1998.
[4] R. DeNicola, G. Ferrari, and R. Pugliese. Types as specifications of access policies. In
J. Vitek and C. Jensen, editors, Secure Internet Programming: Security Issues for Dis-
tributed and Mobile Objects, Lecture Notes in Computer Science. Springer-Verlag, 1999.
[5] C. Fournet, G. Gonthier, J.J. Levy, L. Marganget, and D. Remy. A calculus of mobile
agents. In U. Montanari and V. Sassone, editors, CONCUR: Proceedings of the Interna-
tional Conference on Concurrency Theory, volume 1119 of Lecture Notes in Computer
Science, pages 406-421, Pisa, August 1996. Berlin: Springer-Verlag.
[6] M. Hennessy and J. Riely. Resource access control in systems of mobile agents. In
U. Nestmann and B. Pierce, editors, 3rd International Workshop on High-Level Con-
current Languages (HLCL'98), volume 16(3) of Electronic Notes in Theoretical Com-
puter Science, Nice, September 1998. Elsevier. Available from h t t p : / / w w w . e l s e v i e r .
n l / l o c a t e / e n t c s . Full version available as Sussex CSTR 98/02, 1998. Available from
http://www.cogs.susx.ac.uk/.
[7] X. Leroy and F. Rouaix. Security properties of typed applets. In Conference Record of
the ACM Symposium on Principles of Programming Languages, San Diego, January 1998.
ACM Press.
[8] R. Milner The polyadic 7i-calculus: a tutorial. Technical Report ECS-LFCS-91-180, Labo-
ratory for Foundations of Computer Science, Department of Computer Science, University
of Edinburgh, UK, October 1991. Also in Logic and Algebra of Specification, ed. F. L.
Bauer, W. Brauer and H. Schwichtenberg, Springer-Verlag, 1993.
[9] G. Necula. Proof-carrying code. \n Conference Record of the ACM Symposium on Princi-
ples of Programming Languages, Paris, January 1997. ACM Press.
[10] B. Pierce and D. Sangiorgi. Typing and subtyping for mobile processes. Mathematical
Structures in Computer Science, 6(5):409-454, 1996. Extended abstract in LICS '93.
[11] J. Riely and M. Hennessy. A typed language for distributed mobile processes. In Confer-
ence Record of the 25th ACM SIGPLAN-SIGACT Symposium on Principles of Program-
ming Languages, San Diego, January 1998. ACM Press.
[12] J. Riely and M. Hennessy. Trust and partial typing in open systems of mobile agents.
In Conference Record of the 26th ACM SIGPLAN-SIGACT Symposium on Principles of
Programming Languages, San Antonio, January 1999. ACM Press. Full version available
as Sussex CSTR 98/04, 1998. Available from h t t p : //www. cogs. susx. ac. uk/.
 115

[13] P. Sewell. Global/local subtyping and capability inference for a distributed 7t-calculus. In
Proceedings of the International Colloquium on Automata, Languages and Programming,
volume 1433 of Lecture Notes in Computer Science. Berlin: Springer-Verlag, July 1998.
[14] R. Stata and M. Abadi. A type system for Java bytecode subroutines. In Conference Record
of the ACM Symposium on Principles of Programming Languages, San Diego, January
1998. ACM Press.
 Types as Specifications of Access Policies

Rocco De Nicola, GianLuigi Ferrari, and Rosario Pugliese

Abstract. Mobility is a key concept for network programming; it

has stimulated much research about new programming languages and
paradigms. In the design of programming languages for mobile agents,
i.e. processes which can migrate and execute on new hosts, the integra-
tion of security mechanisms is a major challenge. This paper presents
the security mechanisms of the programming language KLAIM (a Ker-
nel Language for Agents Interaction and Mobility). The language, by
making use of a capability-based type system, provides direct support
for expressing and enforcing policies that control access to resources and
data.

1 Introduction
Most of the difficult issues to face when developing mobile applications running
over a network are related to security. A typical example of security property
is the requirement that only legitimate mobile agents can be granted access to
specific resources, or to specific services. A common solution to this problem is
to provide secure communication channels by means of cryptographic protocols.
These protocols (e.g. SSL, S-HTTP) are designed to provide authentication
facilities of both servers and agents. Another common approach relies on security
architectures which monitor the execution of mobile agents to protect a host from
external attacks to private information.
Recently several researchers have explored the possibility of considering se-
curity issues at the level of language design aiming at embedding protection
mechanisms in the languages. For instance the language Java [4] exploits type
information as a foundation of its security: well-typed Java programs (and the
corresponding verified bytecode) will never compromise the integrity of certain
data.
This idea dates back in time; type systems have been successfully used to
ensure type safety of programs since a long time. Type safety means that every
data will be used consistently with its declaration. However, there has been little
work on exploring and designing type systems for security.
In this paper we discuss the design of the type system for KLAIM (a Kernel
Language for Agents Interaction and Mobility) [14], an experimental program-
ming language specifically designed for programming mobile agents. KLAIM pro-
vides direct support for expressing and enforcing security policies that control
access to resources and data. In particular, the language uses types to protect
resources and data and to estabhsh policies for access control.
The main guidelines for the design of KLAIM and of its type system are:
- KLAIM processes and types are network aware;
 118

— networks and processes are different entities;

— access control policies are network coordination policies.
KLAIM consists of core Linda [17,16,11] with multiple located tuple spaces.
A KLAIM program, called a net, is structured as a collection of nodes. Each node
has a name and consists of a process component and a tuple space component.
Sites are the concrete names of the nodes and are the main linguistic constructs
for providing references to nodes of a net.
Processes may access tuple spaces through explicit naming: operations over
tuple spaces are indexed with their locality. Localities are the symbolic names
of nodes and programmers are not required to know the concrete network refer-
ences, i.e. the precise mapping of localities into sites.
The net primitives are designed to handle all issues related to physical dis-
tribution, scoping and mobility of processes: the visibility of localities, the allo-
cation policies of tuple spaces, the scoping disciplines of mobile agents, etc..
KLAIM types are used to protect resources and data and to establish poli-
cies for access control. Types are abstractions of process behaviours and provide
information about the capabilities of processes, namely the operations processes
intend to perform at a specific locality (downloading/consuming a tuple, pro-
ducing a tuple, activating a process, and creating a new node).
KLAIM capability-based type system encompasses both subtyping and recur-
sively defined types. Subtyping naturally emerges when considering hierarchies
of access rights. Recursive types are used to deal with mobile recursive agents.
The development of KLAIM apphcations proceed in two phases. In the first
phase, processes are programmed while ignoring the precise physical allocations
of tuple spaces and the access rights of processes. An inference system assigns
types to processes, and checks whether processes behave in accordance with their
declared intentions. All rules of the inference system are syntax driven and such
that any process has a minimal type smaller than all its deducible types.
In the second phase, processes are allocated over the nodes of the net. It is
in this phase that access control policies are set up as a result of a coordination
activity among the nodes of the net.
By a mix of both static and dynamic typing KLAIM type system guarantees
that only processes with intentions that match the access rights as granted by
the net coordinators are allowed to proceed.
This paper is a tutorial presentation of the KLAIM approach to mobile agent
security. Its main original contributions are a perspective on types for access
control, some natural extensions of the KLAIM type system and several concrete
programming examples. The interested reader can find additional material in
other publications of our group. The language and the design philosophy under-
lying KLAIM are presented in [14]. The mathematical foundations (decidability
and soundness) of the KLAIM type system can be found in [15] (a preliminary
presentation appeared in [13]). The prototype implementation of the language
is described in [5].
The rest of the paper proceeds as follows. Section 2 illustrates, via simple
programming examples, the basic ideas of the language and of its type system.
Section 3 presents the syntax of processes, types and nets, and outlines the
 119

operational semantics. Sections 4 presents several programming examples and

paradigms that exhibit the flexibility and the expressive power of the language.
The (static) type system which infers types of processes and the network type
checking are formally presented in Section 5. Finally, Section 6 concludes the
paper with a brief discussion of other works and possible extensions. The formal
operational semantics of processes and nets is illustrated in the Appendix.

2 The KLAIM p r o g r a m m i n g model

We begin this section by summarizing the Linda programming model. Then, we

introduce the main features of KLAIM by means of simple examples.
Linda is a coordination language that relies on an asynchronous and associa-
tive communication mechanism based on a shared global environment called Tu-
ple Space (TS). A tuple space is a multiset of tuples, that are sequences of actual
fields (expressions or values) and formal fields (variables). Pattern-matching is
used to select tuples in a TS. Linda provides four main primitives for handling tu-
ples: two (non-blocking) operations add tuples to a TS, two (possibly blocking)
operations read/withdraw tuples from the TS. The Linda asynchronous com-
munication model allows programmers to explicitly control interactions among
processes via shared data and to use the same set of primitives both for data
manipulation and for process synchronization.

2.1 N e t s and Processes

We now introduce the KLAIM primitives without considering typing issues.

KLAIM programs axe structured around the notions of localities, tuples, tuple
spaces, nets and processes.
Localities can be thought of as the symbolic names for sites. Sites are the
addresses (network references) of nodes. The bindings between localities and
sites are stored in the allocation environment of each node of the net. Existence
of a distinguished locality s e l f is assumed that is used by processes to refer to
their current execution site.
Tuples contain information items; their fields can be actual fields (i.e. ex-
pressions, localities, processes) and formal fields (i.e. variables). Syntactically, a
formal field takes the form "\ide", where ide is an identifier. For instance, the
sequence {"foo", 25,! u) is a tuple with three fields. The first two fields are basic
values (a string value and an integer value) while the third field is a variable.
Similarly, {Agent{Itinerary,RetLloc),\ List, Site) is a tuple whose first field is
a process (with two parameters).
Tuple spaces are collections (multisets) of tuples. Pattern-matching is used
to select elements from a tuple space. Two tuples match if they have the same
number of fields and corresponding fields have matching values or variables.
Variables match any value of the same type, and two values match only if they
are identical.
 120

Nets are sets of nodes; each node consists of a site s, an allocation environment
p, a set of running processes P and a tuple space T. We formally model T as
a special kind of process. The allocation environment p constraints network
connectivity in that a node will be able to communicate only with the subset of
the nodes of the network determined by p. Hereafter, we will use
si ::„ Pi I Ti II ... II s„ :>„ Pn \ T„
to denote a net with n nodes.
Processes are the active computational units and may be executed concur-
rently either at the same site or at different sites. Processes can perform five
different basic operations, called actions, that permit reading (writing) from
(in) a tuple space, activating new threads of execution and creating new nodes.
The operation for retrieving information from a node has two variants:
\n{t)@l and read(f)@;. Action \n{t)@l evaluates the tuple t and looks for a
matching tuple t' in the tuple space located at / [l gives the logical address of
the tuple space). Whenever the matching tuple t' is found, it is removed from
the tuple space. The corresponding values of t' are assigned to the variables in
the formal fields of t and the operation terminates; the new bindings are used
by the continuation of the process that has executed in(i)@/. If no matching
tuple is found, the operation is suspended until one becomes available. Action
read(f)@Z differs from in(t)@/ because the tuple t' selected by pattern-matching
is not removed from the tuple space.
The operation for placing information on a node has, again, two variants:
out(t)@/ and eval(P)@/. The operation out(t)@Z adds the tuple resulting from
the evaluation of t to the tuple space located at /. The operation eval(P)@/
spawns a process (whose code is given by P) at the node located at /.
The operation for creating new nodes is n e w l o c ( u i , . . . ,Wn)- It dynamically
creates a set of n different new sites that can only be accessed via locality
variables u i , . . . ,u„.
We now provide some simple examples of KLAIM programs; more advanced
programming examples will be presented later. Hereafter, we assume that the
basic values are i n t e g e r and s t r i n g s .
Our first example illustrates a process that moves along the nodes of a net
with a fixed binding of localities to sites (this somehow corresponds to static
scoping). We consider a net consisting of two sites si and S2- A client process C
is allocated at site Si and a server process S is allocated at site S2. The server
S can accept clients for execution. The client process sends process Q to the
server. This is modelled by the following KLAIM code:

C''=^out(g)@/i.nil
Q '^^ in(7oo", !a;)@self.out(7oo",a; + l)@self .nil
S ''= in(!X)@self .X
The behaviour of the processes above depends on the meaning of /j and self.
It is the allocation environment that establishes the links between localities and
sites. Here, we assume that the allocation environment of site Si, pi, maps s e l f
 121

into si and li into S2, while the allocation environment of site S2, p2, maps
self into S2- Finally, we assume that the tuple spaces located at si and S2 both
contain the tuple ("/oo",l). The following KLAIM program represents the net
discussed above:
si ::,, C|out("/oo",l) || s^ ::,, 5|out("/oo", 1).
The client process C sends process Q for execution at the server node (locality
li is bound to S2 in pi). After the execution of out{Q)@li, the tuple space at
site S2 contains a tuple where the code of process Q is stored. Indeed, it is
Q' '^=^ in{"foo", lx)@si.out{"f 00",x + l)@si.nil
the process stored in the tuple, as the locahties occurring in Q are evaluated using
the environment at site si where the action out has been executed. Hence, when
executed at the server's site the mobile process Q increases tuple "foo" at the
client's site. Fig. 1 gives a pictorial representation of this example.

CLIENT SERVER

\C{"foo",l)\pi I|S(700",1)P2||
I
nil {"foo",l) pi [S{"foo'\l), Q'p2\
i
|nill(7oo",l)pi \Q' {"foo",l)p2\
i
|nil|(7oo",2)pi llnil (7oo'M) P2

Fig. 1. Agent Mobility: Static Scoping

Our second example illustrates how mobile agents migrate with a dynamic
scoping strategy. In this case the client process C is eval(Q)@/i.nil. When
eval{Q)@li is executed, the process Q is spawned at the remote node with-
out evaluating its locahties according to the allocation environment pi. Thus,
the execution of Q will depend only on the allocation environment p2 and Q will
increase tuple "foo" at the server's site. Fig. 2 illustrates this example.

CLIENT SERVER

\\C'{"foo",l)pi\ ||nil|(7oo'M)|p2|

|lnil|(7oo",l)lpi \\Q{"foo",l)p2\

||nil|(7oo",l)|pi ||nil|(7oo",2)p2|l

Fig. 2. Agent Mobility; Dynamic Scoping

 122

2.2 Types for Access Control

KLAIM types provide information about the intentions of processes: download-

ing/consuming tuples, producing tuples, activating processes, and creating new
nodes. We use {r, i, o, e, n} to indicate the set of capabilities, where each symbol
stands for the operation whose name begins with it; r denotes the capability of
executing a read action, i stands for in, o for out, e for eval, and n for newloc.
Semantically, types are functions mapping localities (and locality variables) into
functions from sets of capabihties to types.
Capabilities are used differently by processes and net coordinators. The ca-
pabilities of a locahty I, within a process type, carry information about the
operations the process intends to perform at I. The capabilities specified by a
net coordinator determine the access policy of each site of a net in terms of
access rights and execution privileges. Type checking will guarantee that only
processes whose intentions match the rights granted by the coordinators are
allowed to proceed.
To manifest their access policies processes may annotate localities and local-
ity variables with type specifications. These are pairs of the form (A, 6), where A
is an access list, i.e. a list of bindings from localities to sets of capabilities, and 6
is a type. Access lists will be written as lists of the form [/i t-> TTI , . . . , / „ f-^ 7r„];
(/) denotes the empty list. Processes may use a type specification (A, J) in four
different ways.
- To restrict the actions that a receiving process can perform at (the site
corresponding to) a specific locality when passing locahties around via output
tuple fields of the form I : (A, 6). The special type specification {</>, T) can be
used to state that no further restriction on the access to / is imposed.
- To specify the actions the rest of a reading process intends to perform at the
received site when reading sites via input tuple fields of the form \u : {X,S).
- To specify the access rights of the nodes of the net with respect to newly
created nodes, and the vice versa, when creating new nodes by means of
newloc(u : (A,5)). The special type specification ((j),T) can be used for
stating that a new node inherits the access rights of the creating one.
- To specify, via formal arguments of the form u : (A, (5), the actions that a
process intends to perform at the site denoted by a specific formal argument
of the process definition.
Let us consider a system consisting of a process Server and two identical
processes Client. The server process

Server ''= out(/ : (</>, T))@self .nil

adds a tuple, that contains the locality I, to its local tuple space and evolves
to the terminated process nil; it does not specify any new access policy and
maintains that fixed by the coordinator.
The client process

Client "^^ read(!M : {[self n- {e}], 5p))@/s.eval(P)@u.nil

 123

first accesses the tuple space located at Is to read an address and to assign it
to the locality variable u, then sends process P for execution at u. The type
specification ([self i-> {e}],(5p) expresses that Client is looking for a locality
where it is possible to send for execution a process with type Sp from the site
where Client is running.
Let us now consider a net where Server is allocated at site s and the two,
identical, processes Client are at sites si and S2, where 1$ is bound to s to allow
clients to interact with the server.
Process Client has type

Sc = ls^ {r} t-> -L.

This type states that Client intends to perform a read operation at locality Is-
If the net coordinator grants the following access rights to the sites Si and S2
where Client is willing to execute,
(5j^ = s i-> {r} H-> J. Ss2 = s ^-^ {o} !->• ±
then, after evaluating the localities that appear in Sc according to the allocation
environment of the site where Client is placed, we can see that only the Client
at si has the right of reading tuples at the server's site and, consequently, of
sending processes for execution at I.
The previous example focused on using types as mechanisms for providing
access control policies of sites. The next example, instead, shows how users can
program their own access control pohcies. Consider the user process

P''=^ newloc(!ui : {Xi,6i),\u2 : {X2,S2),\UF : (0,T)).Q

where, for i £ {1,2}, Xi — \U.F ^ Tr^]. The access lists associated to u\ and
U2 mean that all connections to the sites (denoted by) u\ and U2 have to pass
through the site (denoted by) up, which is the only site of the net that has
the rights to perform actions at u\ and U2- The site denoted by up, instead, is
accessible by all the sites that have access to the site where P is running. In
other words, site up is the firewall of the subnet consisting of wi and U2; process
P can program its own firewall policy by sending in execution at up a suitable
firewall manager. External processes will be able to access the firewall only if
its access path is transmitted to them. This can be done by Q above either
via out(uir : (^, T))@/, that allows the receiving process to use up without
any additional type restriction than that fixed at the time up was created, or
via out(ui? : ([/i 1-^ {o},^2 *-> {o}]>T))@Z, that only allows processes running at
(the sites corresponding to) li and I2 to send tuples to up, i.e. to submit service
requests.

3 A Kernel Language for Agents Interaction and Mobility

This section reviews the syntax and the operational semantics of KLAIM. It may
be skipped by readers interested mainly in the programming issues.
 124

3.1 Processes
The syntax of KLAIM terms is given in Table 1; there, P ranges over process
terms, a over actions, t over tuples, / over tuple fields, e over value expressions,
and £ over localities and locality variables. Process and locality variables are
typed whenever they are bound; value variables are kept untyped. We use ~ to
denote a sequence of objects and ( ^ to denote the set of objects in ~ Tuples
are sequences of fields, hence, for a tuple t, {t} will denote the set of fields of t.
Pairs of the form (A, 6) consisting of access lists A and types S will be called type
specifications. Their precise syntax will be introduced in the next section.

p :=nil (null process)

1 a.P (action prefixing)
Pl\P2 (parallel composition)

1 ^ (process variable)
A{P,i,^ (process invocation)

a := out{t)m 1 in{t}m read(t )m 1 eval(P)@£ 1 newloc(M : ( A ^ »

t :=/ 1 /,*
f : = e 1 P \ e •.{X,6) 1 \x 1 !X-.6 \ \u:{X,S)

Table 1. Process Syntax

Actions have been described in Section 2.1. Here we only want to add a few
comments on the action newloc(u : (A,(5)), that dynamically creates a set of
"fresh" nodes together with their access paths u. It is the only action not indexed
with a locality; it is always executed at the current node. For each Ui £ u, \i
specifies the access rights of the nodes of the net with respect to the new node
Ui, symmetrically for S^. The simultaneous creation of a set of new nodes allows
writing mutually recursive type specifications for these nodes.
Variables occurring in KLAIM terms can be bound by prefixes and process
definitions. More precisely, prefixes in{t)@£.. and read(i)@£._ act as binders for
variables in the formal fields of t. Prefix newloc(u : {X,S)).. binds the locality
variables u. Definition A{X : 6,u : {X,d),x) — P is a binder for the variables
{X,u,x\. Hereafter, we shall assume that all bound names in processes are
distinct and shall require that the arguments of eval operations do not contain
free process variables.
We will use the standard notation P[e/x] to indicate the substitution of the
value expression e for the free occurrences of the variable x in P; P[e/x] will
denote the simultaneous substitution of any free occurrence of a; € {x} with the
corresponding e e {e} in P. When substitutions involve locahty variables, e.g.
Hke in P[E/u], they have to be applied also to the type specifications therein.
Notation P[P/X,l/u,e/x\ has the expected meaning.
 125

It is assumed that each process identifier A has a single defining equation

~ —^ —•— ^ def
A{X : S,u : {X,S),x) — P, where process and locahty parameters are expUc-
itly typed. All free (value, process and locality) variables of P are contained in
{X,u,x} and, to guarantee uniqueness of solution of recursive process defini-
tions, all its process variables and identifiers are guarded, i.e. they occur within
the scope of a in/read prefix.
A process is a term without free variables; localities occurring in processes are
considered as constants. In Section 3.3, we will see that they are names whose
meaning is defined (i.e. mapped onto sites) by coordinators.

3.2 Types

Capabilities are elements of {r,i,o,e,n], where each symbol stands for the op-
eration whose name begins with it; r denotes the capability of executing a read
operation, i the capability of executing an in operation, and so on. We use 11,
ranged over by TT, to denote the set of non-empty subsets of [r, i, o, e, n}.
Access lists, ranged over by A, are lists [li y-^ TTi]i=i,...,n, where ii are all
distinct. Semantically, access lists A are partial functions from localities and
locality variables to sets of capabilities (i.e. X{ii) = TTJ). The access list of i,
[ti i-> ni]i=i^,,,^n, specifies the capabilities of £,; (i = 1 , . . . ,n) relatively to £.

S ::=± (empty type)

1T (universal type)
1 ^ l-> TT l-> 5 (arrow type)
1 SuS2 (union type)
U (type variable)
(1V.5 (recursive type)

Table 2. Type Syntax

The syntax of KLAIM types is given in Table 2; there v ranges over type vari-
ables and /x denotes the recursive operator. Hereafter, the following notational
convention will be used: ' W binds stronger than "^", that binds stronger than
",". variables. The type J_ denotes "void", i.e. no intention is declared by the
process, and, semantically, corresponds to the smallest type. Conversely, the type
T denotes the intention of performing any kind of operations and is the greatest
type. A type of the form £ H-> TT i-> (5 describes the intention of performing at t.
those actions allowed by TT, moreover it imposes constraint 6 on the processes
that could possibly be executed at £ (if e ^ TT then (5 is J.). The type 81,62 is the
union of types 61 and ^2; semantically, it is their least upper bound. Recursive
types are used for typing migrating recursive processes.
A type 5 generated from the grammar in Table 2 is such that any recursive
type ^JLV.6' occurring in 5 does not contain v on the left of i->. A consequence
 126

of the requirement that each process variable/identifier in process definitions be

guarded is the fact that we only use recursive types (xu.S whose body S can be
written as 5i,... ,Sn and at least one of the Si has the form i >-^ n i-^ S' (for
some £, TT and 8' such that {r,i} HTT 7^ 0). In the following, we will only consider
types that satisfy this condition; they will be called legal types.
Systems of mutually recursive type equations will be used to define n-tuples
of type. The solution of a system of typed equations is obtained by a standard
iterative technique.
An important feature of KLAIM types is the notion of subtyping. The un-
derlying idea is that if a process P has type 6\ and ^i -< 82 then P could be
considered as having type 82 too. The subtyping relation between 81 and 82 allow
us to say that Pi with type 81 can be safely used in place of P^ with type 82;
but not the vice versa. Subtyping rules and the type inference system will be
presented in Section 5.

3.3 Nets
KLAIM nets are collections of nodes where processes and tuple spaces can be
allocated. A node is a 4-tuple {s,Ps,6s,Ps) where s is a site, P^ is the process
located at s, (5s is the type of s specifying the access control policy of s, and ps is
the allocation environment of s, i.e. a (partial) function from localities to sites.
We write s ::*^ Ps to denote the node {s,Ps,Ss,ps)-
Hereafter, £ will denote the set of environments, (p the empty environment,
and {s/l} the environment that maps the locality I on the site s. We will use
£{p} to denote p{E), if p{£) is defined, and £, otherwise; moreover, p[s/i] will
denote the environment p' such that p'{£) = s and p'(i') = p{i') for £' ^ £.
To specify the mutual access policies of a set of nodes, hence to consistently
assign types to sites/nodes, we make use of a partial function A that, for each
site s, describes the access rights of s on the other sites. Types of nodes have
the same syntax of types of processes. However, strictly speaking, the formers
cannot be generated by the grammar given in Table 2, since we required £ to
stand for localities and locality variables. For types of nodes, we let £ range over
sites.

Definition 1. A KLAIM net is a pair Ns : A where Ns is a finite multiset of

nodes over the set of sites S, and A : S -^ {S -^ U) is a partial function.
Function A induces a system of n type equations in the variables {I'si | s, S 5}
where, if { s } , . . . , s^.} = dom{A{si)), the equation of site Si is
Us, =s\ ^A{si){s{) ^Usi^,---,si>^A{si){si)^i^si,
and the following conditions hold:
1. for any node s ::^"^ P, Ps(self) = s, image{ps) C S and 8s is the component
of the solution of the system of type equations induced by A corresponding
to s;
2. for any pair of nodes s ::*^ P and s' ::p'' P', s = s' implies ps = ps' and
 127

Condition 1. expresses that the local allocation environment of a node maps

localities on the sites of the net; in particular, s e l f is mapped on the name
(site) of the node. Moreover, the type of the node is determined by the function
A associated to the net. The idea underlying condition 2. is that the nodes of
a net over the same site do have the same allocation environment and type as
well.
Nets will also be written according to the syntax given in Table 3.

T a b l e 3 . Net Syntax

When developing KLAIM appUcations, type checking (that shall be presented

in Section 5) will be used to guarantee that the types of the processes in the net
agree with the access rights of the sites where they are located. More specifically,
types of processes are checked against those fixed for the nodes of the net by the
net coordinator, while taking into account where each process has been located.
To compare process types, that are functions mapping localities into functions
from sets of capabilities to process types, with node types, that are functions
mapping sites into functions from sets of capabilities to node types, localities
have to be mapped into sites by using site allocation environments, i.e. localities
must be interpreted.

Definition 2. The type interpretation function associated to a net Ns, ©NS •

S —> £ is defined as follows: for all s e S, 0NS{S) = Ps if S ::*^ P G Ns, for
some Ss and P.

The type interpretation function is well-defined because A^5 enjoys conditions

1. and 2. of Definition 1.

3.4 Operational Semantics

We describe the operational behaviour of nets using a reduction semantics. A

reduction semantics is centered around the notions of structural congruence and
reduction relation. They are formally presented in the Appendix, here we provide
some comments to enable the reader to understand the examples.
The structural congruence provides a convenient way of rearranging nodes.
The structural laws express that || is commutative and associative, and that it
is always possible to distribute the processes located onto the same node over
clones of that node.
The reduction relation specifies the basic computational steps. The execution
of an out operation adds a tuple to a tuple space. The local allocation environ-
ment is used both to determine the site where the tuple must be placed and to
 128

evaluate the tuple. If the tuple contains a field with a process, the corresponding
field of the evaluated tuple contains the process resulting from the evaluation
of the used localities by using the local allocation environment. The fact that
processes in tuples are transmitted after the interpretation of their localities
corresponds to having a static scoping discipline for the generation of tuples.
A dynamic scoping strategy is adopted for the eval operation. In this case the
localities of the spawned process are not interpreted using the local allocation
environment. Processes not closures are transmitted and their execution can be
influenced by the remote allocation environment.
A process can perform an in action by synchronizing with (a process rep-
resenting) a matching tuple et. To match the two candidate tuples one has to
consider the site where the operation is executed and the type interpretation
function of the net. The result of the execution of an in action is that tuple et
is withdrawn and its values are used to determine the values of the variables in
the input tuple to be used by the continuation. Action read behaves similarly
to in but leaves the matched tuple et in the tuple space.
The use of newloc leads to the creation of new nodes and modifies the
topology and the types of the net. The allocation environment of the new nodes
is derived from that of the creating node with the obvious update for the s e l f
locality. The type specifications are exploited to generate the new types. The
types of the nodes of the net are "extended" by adding the rights of the existing
nodes over the new ones and the rights of the new nodes over the existing ones.
Access lists are used to enrich the rights of the existing nodes, while types are
used to determine the rights of the new ones.
Pattern-matching is extensively used in the reduction semantics. To select
(from a tuple space) a tuple containing actual fields with processes inside the
pattern-matching operation checks the types of processes to ensure that they
satisfy the type constraints specified by the programmer. Hence, process codes
are checked before being downloaded. In other words, the pattern-matching
operation performs a run-time type checking of incoming codes.

4 Examples and Programming Paradigms

In this section we show how to program in KLAIM by means of some examples.
We illustrate a mobile agent for information retrieval, some basic paradigms for
mobile code applications and the implementation of a few access restrictions
over nets.

4.1 Distributed Information Retrieval

Distributed information retrieval applications gather information from a set of
sources distributed over a network; the nodes to be visited may be determined
either statically or dynamically. Here, we present a KLAIM solution of a dis-
tributed information retrieval problem and discuss some of the access control
issues there involved.
 129

Consider a user process t h a t needs information about a d a t a of which he only

has a key represented, say, by "item", to be used for searching in a d a t a b a s e dis-
tributed over the network. In our solution, the distributed database is modelled
by located tuple spaces and it is assumed t h a t each local d a t a b a s e (a tuple space)
reachable from luem (the starting point of the search which is known by the user
process) contains either a tuple of the form i^'item",v) with the required infor-
mation V, or a tuple of the form H'item",snext), with the site of the next node
to visit.
T h e user process UP asks for the execution at litem of the mobile gatherer
agent, G, and then retrieves the result of the search at the (new) node UR^

UP'^= newloc(wfi : ( [ s e l f i-> {i},litem M- {o},li H- {o},l2 >^ {o}],-L)).

Agent G travels across the nodes looking for tuples containing information
associated t o the search key t h a t is passed to it as parameter. G is p r o g r a m m e d
in K L A I M as follows

G{x,u : ( [ s e l f H^ {O}],±)) '^= read(a:;, !2/)@self .out(j/)@u.nil

I read(a:, !u' : ( [ s e l f v^ {e}], J ' ) ) @ s e l f .eval(G(a;, w))@u'.nil.

Assume t h a t P has type 5p in a context t h a t binds all free variables in P.

Then t h e types of UP and G{x, u : ( [ s e l f >-¥ {o}], J.)) are:
SUP = 6p, s e l f ^ {n} ^ ±, ktem i-^ {e} >-^ htem H-^ {r} t-> ±
SG = s e l f !-)• {r} h-> ±,M H-> {o} >->• -L.
Notice t h a t t h e type of a closed process does not contain locality variables.
The operations the process intends to perform at addresses denoted by locality
variables are checked against the type specifications for those variables during
the inference and are not reported in the type of the process.
As an example of type constraint checked in the type inference, we have t h a t
type S', t h a t we deliberately have not specified, must be greater t h a n SQ, i.e. at
least all the needed rights must be asked for.
Let us now see how types can be used to enforce access control policies by
considering a net with three sites s, si and S2, with types Ss, Sgi and 6,^. Assume
t h a t si contains the tuple {"item,", S2), S2 contains the tuple {"item",v), UP is
allocated at site s, Ps{litem) = s i , Ps{h) = si and ps{l2) = S2- T h e K L A I M code
of the net is:

s ::^"^ UP II si :•.%] out{"item", S2) \\ S2 ::''4 out{"item", v).

Checking whether the net is well-typed requires determining the interpre-
tation of SUP with respect to the allocation environments and comparing the
resulting type with the access control policy given by Ss- T h e interpretation of
SUP is

^ Here, localities h and Z2 are used to represent the addresses of two nodes different
from the node where user process is allocated. Processes located at these nodes are
allowed to perform only output actions at the node named by UR. The use of Zi and
I2 could be avoided by introducing in the type specifications a distinguished locality
others to denote any node other than those explicitly mentioned.
 130

[Sp], s M- {n} !->• _L, si H^ {e} M- si i-> {r} i->- 1

where [Sp] denotes the interpretation ofSp. To ensure well-typedness, we should
have that [dup] :^ Sg which asks for:
[Sp] <Ss, s y-^ {n} i-^ ± < Ss and si H-> {e} H- (si i->- {r} i-^ ±) <Ss.
The type S' of the variable u' (used by the agent G when reading continuation
sites) is used statically (by the type inference system) to check whether So :< S',
and dynamically (by the pattern-matching mechanism) to select the appropriate
tuple. For instance, the gatherer agent G can travel from si to 5-2 only if the tuple
{"iteTn",S2) can be selected. This amounts to requiring that the interpretation
of S' should not be greater than Ss2 •

4.2 Code Mobility

Mobile Code Applications are applications running over a network whose distinc-
tive feature is the exploitation of forms of "mobility". According to the classi-
fication proposed in [12], we can single out three paradigms that, together the
traditional client-server paradigm, are largely used to build mobile code appli-
cations. These paradigms are:
— Code-On-Demand;
— Remote Evaluation;
— Mobile Agent.
Code-on-Demand A component of an application running over a network on a
given node, can dynamically download some code from a remote node and link
it to perform a given task.
To download code that respects certain type constraints S and is stored in
a tuple at a remote tuple space /, action read(!X : S)@l.X can be used. The
downloaded code is checked for access violations at run-time by the pattern-
matching mechanism. The typing rules for read and in are designed in such a
way that any downloaded code whose type is a subtype of S will not violate type
correctness.

Remote Evaluation Any component of a networking application can invoke ser-

vices from other components by transmitting both the data needed to perform
the service and the code that describes how to perform the service.
To transmit both code P and data v at the locality / of the server, action
out{in{\y)@l.A{y),v)@l, where A{x) — P, can be used. Here, we assume that
the server adopts the following (code-on-demand) protocol
inQ.X : J, !a;)@self.out(a;)@self.X.
To prevent "damages" from P , the server may mount and execute code P only
if [Sp], the type of the code P when interpreted at the server's site, is such that
[Sp] •< [S], where [S] is the interpretation of S at the server's site.
Type S may give only minimal access permissions on the server's site, for
instance, only the capability of reading some resources and giving back the results
of the execution. In this case, [S] is of the form
 131

S l-> {r} 1-4 ±, So •->• {o} 1-4 ±, Si 1-4 {o} 1-4 ±, . . . S„ 1-4 {o} H4 ± ,

where Si, i = 0,.. .,n are the sites with the rights of invoking server's faciUties,
and s is the server's site.
Notice, however, that this does not prevent P from visiting other sites. In
particular, agent P may be programmed in such a way that after having per-
formed the required elaboration, it transmits code Q at the locality /j (the logical
name of site Sj):

P '^= read(!a:)@/.out(op(a;,y))@self .out(g)@/i.nil.

It is immediate to see that, if we assume that the client is allocated at site SQ
and that P has type 6p = I >-^ {r} i-4 J., s e l f >-4 {o} i-4 -L, Zj 1-4 {0} ^4 J., then
[Sp] = s i-> {r} H4 J-, So 1-4 {0} 1-4 -L, Si 1-4 {0} f-4 J.. Hence, [Sp] < [6] However,
code Q is only stored in the tuple space at SJ: no new thread of execution is
activated at site Sj. Before being executed code Q must be read and verified
(dynamic type checking). Therefore, process P cannot activate a Trojan horse
at the remote site Sj.

Mobile Agents A process (i.e. a program and an associated state of execution)

on a given node of a network can migrate to a different node where it continues
its execution from the current state.
If we want to adopt a dynamic scoping discipline for a mobile agent A that
migrates to locality / for execution, the paradigm can be programmed using
action eval(A)@/. If a static scoping discipline is preferred, then action out(A)@/
can be used instead (of course a code-on-demand protocol is assumed to pick
up and execute ^4).
With dynamic scoping, the access rights of the mobile agent A are controlled
statically when the code of action eval(y4)@Z is encountered. With static scoping,
access rights of A are controlled dynamically by the pattern-matching when the
tuple containing the code of agent A is picked up from the tuple space.

4.3 Protecting the net

We now show how types can be used at the level of net coordinators to control
accesses. In other words, we show how types can be used to set up the security
architecture of KLAIM nets.

Restricting Interactions KLAIM action primitives operate on the whole net: oper-
ations on tuple spaces are not forced to happen locally at their current site. Prom
the point of view of access control policies, communications among different sites
of the net (i.e. remote communications) could be controlled and regulated.
To force a process running on a certain site s to access only local tuples it
suffices to constrain the type 6s of the site in such a way that for no s' ^ s it holds
that s' 1-4 {r} H4 ± :^ i5s. Hence, a process P allocated on s performing remote
read/in operations violate the access rights. To access tuples at a remote tuple
space, a well-typed process must first move (if it has the required rights) to the
 132

remote site. Also output actions can be forced to be local; it suffices requiring
that for no s' ^ s it holds that s' i-> {o} y-^ ± < dg.

Firewall We now show how to specify a firewall by means of suitable typing.

The idea is that the type of the site where the firewall is allocated specifies the
access policy of network services. Assume that the firewall is set up at site s to
protect sites s i , . . . , s„j from sites s^,..., s^.
A first requirement is that processes located at site s'- cannot directly interact
and ask for services at any site Sj. To this purpose we impose that for each site
s',, for no 7r it holds that Sj (-> TT H->- _L ^ ^ /. Instead, type 6s' may have subtypes
of the form s i-> {o} »-> i.. This means that each site s'j may only send service
requests to the firewall (here we assume that the service request is stored in a
tuple). In other words all connections from sites s ' j , . . . , s^ to sites si,... ,Sm
have to pass through site s, the firewall.
def
For instance, process Pi = out(eval((3)@/s;)@/s.nil complies with the ac-
def
cess rights of a site s'^, while process P2 = eval(Q)@isi.nil violates the access
rights of site s'^..
The firewall can be programmed to handle the requests according to certain
policies. Typically, the firewall handles a service request by a process of the form
reaid{\Request : ^).(RequestHandler)
where type 6 specifies the access policy that the service request must satisfy. For
instance, if
6 - si ^ {i} >-^ ±, S2 i-> {e} i-J- (s2 !->• {r} H-J- _L, S'^. H4- {0} i->- ±),

then the service requests eval(read(!a;)@self .out(op(x))@Zs' .nil)@/s2-iiil ^^^

read(!a;)@isi.nil both satisfy the type requirements and will be accepted. The
request read(!a;)@/sj .out(op(a;))@/s' .nil violates the type constraints and will
be rejected.
One can easily see that, by adding standard types for tuples (basically record
types), it is possible to specify more refined policies to handle service requests.
For instance, one can fix a policy where the read operations are constrained to
get from the tuple space only 'plain' data tuples, without localities or process
codes.
As we outlined in Section 2.2, a user can program its own firewall. Let us
consider, for instance, the following user process:
def
P = newloc(!ui : {AI,JI),!M2 : (A2,(52),.. .!u„ : (A„,<5„), luir : {XF,6F))-Q,
where, for alH £ { 1 , . . . , n } , Ai = [up >-> TT^]. Process P creates a subnet whose
sites are associated to locality variables u i , . . . ,u„; the type specifications {\i,6i)
ensures that the subnet can be reached only through site UF (the firewall). The
user process can specify in the type specification {\F,6F) the access policy to
the firewall. For instance, setting Xp = s e l f i-> TT means that the firewall site
can be accessed only from the site where the user process is running.
 133

Fares and Tickets A primary access control policy consists of controlling the
route of a mobile agent traveling in the net. For instance, if one has to configure
a set of sites with new software, a mobile agent can be programmed to travel
among the sites to install the new release of the software.
If the starting site of the trip is site SQ and sites si, S2,..., Sn are visited
before coming back to the starting site, then the following equalities specify the
type 6T (the access rights) of the trip:
do = si \-^ {e} I-)- (5i,(5
(5i = S2 M- {e} H-)- 82,5

Sn-i == Sn I-*- {e} *->• S',S

(5' = So M- {0} i-¥ 1

The idea is that at each site type <5 specifies the allowed operations (e.g. installing
the new release of a software package); the remaining type information specifies
the structure of the trip (which is the next site of the trip). At the last site of
the trip the agent has the rights of returning to the original site the results of
the trip (e.g. the notification that the installation was successful).
The type discussed above can be properly interpreted as the fare of the trip:
an agent A can perform the trip provided that its type 5A matches the fare:
formally, 6A when interpreted at site SQ is a subtype of ST- In this case the agent
A has the ticket for the trip. Notice that this ensures that a mahcious agent
cannot modify the itinerary of the trip to visit other sites different from those
listed in its ticket.

5 A Capability-Based Type System

This section formally presents the KLAIM type system that generalizes the one
of [15] and differs from it in three respects. Here, in addition to types, type
specifications are used. Moreover, type inference is parameterized by the locality
where the process is supposed to be located. Finally, derived types are completely
independent of the names of the locality variables occurring in processes.

Definition 3. A type is in canonical form if it is generated by the following

grammar
6 ::= J. 1 T 1 (/>!,..., 0 „ 1 IJ,V.{(l)i,...,(p„) (n>l)
(j) ::= u 1 ^ 1-^ TT (->. (5
where (j)i,... , 0 „ and S satisfy the following constraints:
all bound t y p e variables are distinct;
all union types (pi,.,. ,(j)k are such t h a t , for all i,j {I < i,j < k and i ^ j),
if (pi and (pj are variables then (pi ^ (pj, otherwise L{(pi) n L{(pj) = 0, where
L{S), the set of shallow localities of 5, is defined inductively on the syntax of
6 as follows:
 134

L{±) = L(T) = L{u) =% L{i^n^S) = {1}

L{di,d2) = L(Ji) U L{d2) L{^iu.6) = L{6);
— all /x-types fifS' are such that u occurs in 6' and it is always "guarded",
i.e. on the right of an even number of H->'S; moreover, li S' = (f>i,... ,(})n then
(f>i ^ V for a l H {1 < i < k).

Roughly speaking, in a canonical form ijii,..., 0„ (n > 1) each component <^i

(1 < i < n) is either a variable or an arrow type, with the additional constraint
that all shallow localities be different. The condition about shallow localities
allows us to determine the capability and the type associated to a given locality
within a canonical form.

Equalities on union types

{l)Su 62 = 62 61
(2) {61,62),63 = 61,(62,63)
(3)6,6 = 6
(A) 1,6 = 6
(5)T,<5 = T
( 6 ) £ l-> TTl t-^ J i , £ i-> 7r2 !->• 62 = f !->• TTl U 7r2 i-> (-51 62)

Equalities on recursive types

(7) tiVi.5i,nv-2 .62 = iJ,u.(6i[i/li/i\ 62[v/>^2]) (a-renaming)

(8) 6[nu.6/u] --= iii'.6 (folding/unfolding)
(9) 6[6i/i^] = <h a n d 6[62li>] = 62 imply 61 = ^2 (contraction)

Table 4. Type Equality

In [15] an algorithm is given for reducing (legal) types to canonical forms.

The algorithm relies on the type equality relation, = , the least congruence that
satisfies the rules in Table 4. Rule (9) holds only when S is contractive in f,
that is when u occurs in S only under 1-4. Decidability of = has been proved
in [15]. From now onward we will only consider types in canonical forms (and
their unfoldings). With abuse of notation, we will continue using S also to refer
to canonical forms.
To define the subtyping relation, •<, we start by introducing an ordering
between capabilities, namely a hierarchy over access rights. The chosen ordering
relies on the assumptions that if a process is able to perform an in operation then
it is also able to perform a read and that if a process has a set of capabilities n
then it posses any subset of TT. Hence, equality of capabilities must be intended
as standard set-equality but {i,r} = {i}-

Definition 4. The ordering relation on capabilities, C.^, is the least reflexive

and transitive relation induced by the rules in Table 5.
 135

r~ f r- '
TTl C 7r2
T^i L j j ^ 1 '^2 !=n "'a
{i} ^n {r}
"•a C „ TTi (TTI U Hi) Q„ (TT'I U TTJ)

Table 5. Capability Ordering Rules

Capabilities only consider the intentions of processes relatively to fixed lo-
calities, while types consider also the intentions after migration. The subtype
relation introduces a hierarchy over intentions.

A.<6 (axl)

(axT)

6^5'
5^6' (eq)

S, ;< S'l or or Si -< Si for all 1 < i < n

Si,. . . ,Sn <S'i,...,5'k (u/u)

^2 E „ T i , Si :< S2

£ (->• TTl 1-4 J l :< e 1-^ TT'Z ^-^ S2

e>-^n^-^S :<S'[iJ.v.S'/iy]
i>-^ n 1-^ S :< nu.S'

HU.S' •< (. i-^ IT i-^ S

Sx\v/v-i] -< SiSyjv-iX

where i^ is a fresh variable {\xj{i)
iW\ .S\ -< [W2 M

Table 6. Subtyping Rules

Definition 5. The subtyping relation over canonical forms and unfoldings of

canonical forms, :<, is the least relation closed under the rules in Table 6.

Decidabihty of subtyping has been proved in [15].

5.1 Typing Processes

Type contexts, F are functions mapping process variables and identifiers into
types, and locality variables into type specifications. Type contexts are written
as sequences of assignments X^ : (5i,... ,X„ : (5„,ui : {\i,5[),... ,Um'• {^m,^'m)-
The special symbol (j) is used to denote the empty context.
Before presenting the judgments of the type inference system we introduce
some useful notations.
 136

First, we equip type specifications with the (postfixed) projection operations

4-1 and 4-2 defined as follows: {X,S) 4-i = A and {X,S) 4-2 = S.
Notation 6@£, defined below by induction on the syntax of S, denotes the
partial evaluation of type 6 which replaces self with the locality i.

i^TTM- {6'm) if£' = self

(£ ^ T H^ (5 )(L. - ^ ^, ^ ^ ^ ^^,^^,^ otherwise;

- {nu.6')m = fiiy.is'm).
Hereafter, we write 6@i to denote a canonical form of type d@£. We also write
£'m to denote £ ii t = self, £' otherwise.
Given a type context F, we write r[S/X] to denote either the extension of
r with the assignment X : 5 (when X is unbound in F), or the updating of F
that binds X to S {F[{X,6)/u] has a similar meaning). The auxiliary function
updatef, indexed by the locality £ where a process is located when the function is
invoked, behaves hke the identity function for all fields but \X : 6 and ! w : (A, ^).
In the former case, the behaviour of the function is obvious. In the latter case,
the type specification assigned to u is obtained by replacing self with ^ in A
and by partial evaluating 6 at u. Formally, it is defined by:

{
updatef {update^{F,f),t) \i t = f,t,

F[S/X] iit = lX:S,

r[{A[£/self],(5@u)/u] li t = \u : {X,6),
F otherwise.
To check that the type specifications occurring in a tuple t are all consistent at
£, we require that, for each u : (A„,(5„) £ {t}, updaie,((/>, i)|-^ (5,i holds, where
r | - y 5 is the least predicate closed under the following rules:
r\- ± r\^T r\-f-u
e'edom{r) {r{e')i,){e) n„ n 5'^r(£')4-2 e'^domir) r|-^<5'
r\-^ <5i r | - p 6-2 r\-f- &
r|-p^i,52 r\-i-iiv.5
We will always assume that update^{F,t) is defined only if for each u : (Au,(5„) €
{t}, update^{(f>,t)\Y- Su holds.
Notation update^{F,u : {\,5u)) has the expected meaning, and is defined
only if the type specifications there involved are all consistent. This amounts to
saying that, for each u € [u], updatef{4>,u : {A, (5„))|-^ (5„ holds.
Given a type context F, a locality £, a set of locality variables U such that
dom{F) C U and a type S, we write F\Y- S\y to denote the type obtained by
deleting in 6 the subterms relative to the variables in U while checking that 6
satisfies the corresponding type specifications stored in F. Formally, Tf-j- 6\^
is defined by induction on the syntax of S by the following rules:
 137

r^i^^=± r^TN„=T r^^^^=^

r\-p{5i,62)\^ =S'i,6'2 r\-^ {fj.u.6)\^ = )j.v.5'

The type judgments for processes take the form r\j- P '• ^ where J" is a
type context providing the type of free locaUty and process variables and that
of process identifiers in P. The judgment r\-j- P '• ^ asserts that, within the
context r, the intentions of P when located at (. are those specified in 6.
Type judgments are inferred by using the rules in Table 7. There, given a
tuple t, we write lv{t) to denote the set of locality variables occurring in t, i.e.
the set {u \ 3A,6 : u: {\,5) e {t}}. Moreover, in the conclusion of the inference
rules, we shall write {61,82) to denote the canonical form of the type {5i,82)
obtained by applying the rules in Table 4. For instance, we shall write that from
r\Y- P : i' ^ IT <-^ 6 we derive r\-j- out(i)@£'.P : £' H- TT H-J. 5,£' i-> {0} H^. ±
but actually we mean F^j- out{t)@i'.P : £' i-> TT U {0} M- 6. We now comment
on the inference rules.
The simplest process (the null process nil) has no intentions at all. The type
of a process variable (or identifier) is completely determined by the type context,
r, and by the locality, £, where the process is supposed to be at. Definedness of
r{X) is guaranteed by the fact that processes are closed terms.
The typing rule for out states that the type of out{t)@£'.P extends the
capabilities of P at i'@i with o. Since out is not a binder, P is typed within the
same context as out(t)@£'.P.
The typing rule for read first derives the type (5 of P in a context updated
with the type assignments for the process and the locality variables bound by
read. Then it removes from S the subterms relative to the locality variables
that are in update^ {P, t) and not in P while checking that 5 satisfies the type
specifications that correspond to these variables. Finally, it extends the obtained
type at £'@£ with capability r. The typing rule for in behaves similarly.
The typing rule for eval extends the type of P at i'@t with e and records
that the operations of P at t'@l have to be extended with those (given by 6') of
the spawned process Q at £'@£.
The typing rule for newloc first derives the type (5 of P in a context updated
with the assignments of the type specifications declared for the locality variables.
Then, it discharges the type information associated to the locality variables and
checks that 5 satisfies the type specifications that correspond to these variables.
Finally, it extends the obtained type at £ with capability n.
The typing rule for parallel composition states that the intentions of the
composed process are the union of those of the components, while the binding
context is left unchanged.
The typing rule for process definition updates the type context with the
types declared for the process variables occurring as parameters in the definition
 138

r(-j- nil : J.
r|-p X • r{x)@i r^r A •. r{A)m
rirP'-S
rf-p out(t)@/.p: {s,{e'@e) n- {o} ^^ ±)
updatef{r,t)\-j- P •S updatef{r,t)\-f- <5N,„,,, = S'
r\-p read{t)m'.P : {5', {£'©£) i-> {r] H-> L)
updatef{r,t)[j-P :5 updatei{r,t)\-j- S\,^^ = S'
r\j- in{t)m'.p: {6', {e'm) *->. {i} ^ ±)
r\^P:5 r h p ^ Q : 6'
r\-f- eval(Q)@/.P : {5, {£'©£) i-> {e} H4 6')

updatef{r,u : {X,5u))\-(-P •S iipdatej(r", u : (A, (5M))|-J-5\ =5

r | - j - n e w l o c ( S : ( A X ) ) - P : (<5', ^ M- {n} H4 ± )

r\^P\Q:{S,,5-2)

r[X:'hc][A:S]\^^P:S' S^S' uprfate„,,(<^, S: ( A % ) ) h ^ ^ ^ J ' N ^ ^ ^ = 5"

r\-^A:S
' self

^ b n r ^ '• * ^ h r ^ ' ^ ^i and .5i ^ (<5xi@^) for all P G {P}

r\^A(p,e,e}: {5@e)[i/^
Table 7. KLAIM Type Inference Rules (we assume that A{X •.5x,u:{\,5u),x) = P)

and with the binding between the process identifier A and a (possibly recur-
sive) candidate type 6. The resulting context is exploited to infer (up to type
equality) the type S for P. A second type inference is triggered by the rule, start-
ing from a type context that additionally contains the bindings for the locality
variables occurring as parameters in the process definition. This last inference
checks whether the intentions of process P comply with the type specifications
of its locality variables. To make types of process identifiers independent of the
localities where the identifiers are invoked, it is assumed that the latters are
always invoked at self.
The last typing rule is the rule for process invocation. First, it determines the
type of the process identifier and those of the process arguments. Then, it checks
whether the type inferred for any process argument agrees with that obtained
by partially evaluating at i the type of the corresponding formal parameter. No
requirement is imposed on the other arguments. The inferred type states that,
once we interpret the type of A starting from i (the actual locality where the
invocation takes place), A{P, t, e) intends to perform at I the same operations of
A at u. Indeed, the locality variables occurring as parameters in the definition of
 139

the process may occur in the type inferred for the process identifier. Soundness
of the appUcation of [t/u\ to S follows from the assumption that all bound names
in the definition of A are distinct.
In [15], it is proved decidability and derivability of a minimal type for any
typable process. The type is called minimal because all the other deducible types
are greater than it. The main impact of the existence of a minimal type is that
the type inference system is decidable. The same results holds for the type system
presented in this paper.

Theorem 1. li r\j- P : S' then there exists a minimal type S such that r\-j-P :
S a.nd6^ 6" for all 6" such that T ^ ^ P:6".

Corollary 1. For any process P, the existence of a type 6 such that <j)\-j- P '• S
is decidable.

5.2 Typing Nets

In this section, we introduce the notion of well-typed net of processes.

Definition 6. Given a type interpretation function Q^Vs, a node s Q S and a

type S, then the interpretation S{0}Vs }« of (5 in s is a canonical form of the type
defined inductively as follows:
- ±{6»iVs}s = -L, '^{9NS]S = T, iy{9N^}s = v;

\ / 1 jvs J s 1^ ^ ^_^ ^ |_^ (5' otherwise;

- {iJiv.6'){GNs]s = HP.5'{9NS]S-

Hence, localities that cannot be mapped to sites of the net are left unchanged.

Definition 7. A net Ns is well-typed if for any node s r:^^ P, there exists 6'
such that (f>\-^ P : 6' and if J is a minimal type for P then S{0j^g}s •< Ss-

The weil-typedness condition requires that processes cannot use a locality if

this is not known in the allocation environment of the node.
The KLAIM type system enjoys two key properties: subject reduction and type
safety. The former property guarantees that well-typedness is an invariant of the
operational semantics. The latter guarantees that well-typed nets are free from
run-time errors. Such errors are generated when processes attempt to execute
actions that are not allowed by the capabilities they own. The two properties
together amount to saying that well-typed nets never encounter run-time errors
due to misuse of access rights. In other words, the errors that the KLAIM type
system precludes are those due to access right violations. We refer to [15] for the
statements and the proofs of the two properties.
 140

6 Concluding R e m a r k s
We have developed a type system which formahzes access control restrictions of
programs written in KLAIM. Type information is used to specify access rights and
execution privileges, and to detect violations of these policies. The implementa-
tion of the type inference system for X - K L A I M (the prototype implementation
of KLAIM) is in progress; it will help us also to assess our design choices.
We plan to extend the type system by introducing types for tuples (record
types), notions of multi-level security (by structuring localities into levels of
security) and public or shared keys to model dynamic transmission of access
rights. Ideas could also be borrowed from the spi-calculus [2], a concurrent cal-
culus obtained by adding public-key encryption primitives to the 7r-calculus
[21], and from the SLam calculus [18], another calculus where information about
direct/indirect producers and consumers are associated to data.
Another direction for future research is considering "open" systems. In fact,
our type system can safely deal with new processes landing on existing nodes,
but it does not consider partially specified nets. An enrichment of types is then
needed to specify the permissions granted to "unspecified" sites. Then, the choice
has to be faced whether this enrichment should be specified at the level of single
nodes or at the level of nets. Interfacing nets would naturally fit with extensions
of our framework to hierarchical nets that would be beneficial also for more
structured access controls. An alternative approach to deal with open systems
could also be that of relaxing the static type checking phase by not requir-
ing well-typedness of the whole net (this corresponds to the fact that only the
typed sites can be trusted) while increasing the run-time type checking phase,
e.g. agents migrating from untyped (i.e. untrusted) sites must be dynamically
typechecked. This is the approach followed in [24].
Type systems have been used also for other calculi of mobile processes.
Among those reminiscent of ours, although not addressing security issues, we
mention the work of Pierce and Sangiorgi [23]. They develop a type system for
the TT-calculus using channels types to specify whether channels are used to read
or to write. This type system has been extended in [20] by associating multi-
plicities to types for stating the number of times each channel can be used. The
type system of [23] has been also generalized by Sewell [25] to capture locality of
channel names and by Boreale and Sangiorgi [7] to trimmer bisimulation proofs.
Only recently attempts have been made to characterize security properties
in terms of formal type systems. A type system for the spi-calculus has been
developed by Abadi [1] to guarantee secrecy of cryptographic protocols. Abadi
and Stata [3] have used type rules to specify and verify correctness of Java
Bytecode Verifier. Hennessy and Riely [19] have introduced a type system for
the language DTT, a distributed variant of vr-calculus for controlling the use of
resources in nets. This work is similar to ours (types are abstraction of pro-
cess behaviours and access rights violations are type errors), but the technical
developments are quite different; resources are channels and types describe per-
missions to use channels. Moreover, access rights are fixed irrespectively of the
localities where processes themselves are executed. In [24], the type system of
 141

[19] has been improved for considering nets where sites can also put up malicious
agents that do not respect the rules on the use of resources. Cardelli and Gordon
[10] have introduced a type system for mobile ambients [9] that controls the type
of the values exchanged among administrative domains (ambients) so that the
communication of values cannot cause run-time faults. Volpano and Smith have
developed type systems to ensure secure information flow (noninterference) for
both a sequential procedural language [27] and for a multithreaded imperative
language [28]. Boudol [8] has used types to abstract from terms the possible
sequences of interactions and the resources used by processes. Necula [22] has
introduced an approach to ensure correctness of mobile code with respect to a
fixed safety policy, where code producers provide the code with a proof of cor-
rectness that code consumers check before allowing the code to execute. Vitek
and Castagna [26] have proposed a language-based approach, relying on power-
ful mobility and protection primitives, rather than on type systems, for secure
Internet programming. Bodei, Degano, Nielson and Nielson [6] have proposed
an alternative approach for the analysis of security and of information flow, that
relies on static analysis techniques.

Acknowledgments
We would like to thank Betti Venneri for her contribution to the development of
the KLAIM type system. We are also grateful to Lorenzo Bettini, Michele Boreale
and Michele Loreti for general discussions about KLAIM.
This work has been partially supported by Esprit Working Groups CON-
FER2 and COORDINA, and by CNR: Progetti "Metodologie e Strumenti di
Analisi, Verifica e Validazione per Sistemi Software AfEdabili" e "Modelli e
Metodi per la Matematica e I'lngegneria".

A Operational Semantics

In this section, we coerce i to denote localities, locahty variables and sites,

and assume that allocation environments are extended to sites; when applied to
sites, allocation environments act as the identity function. Tuples are modelled as
processes and the KLAIM syntax is extended with processes of the form out(ei)
for denoting evaluated tuples (referred to as et).
The KLAIM operational semantics uses an evaluation function for tuples,
T | - ] ^ , that is inductively defined over the syntax of tuples by the rules in
Table 8, where we use Sfej to denote the value of the closed expression e.
There are only two non-trivial cases, namely Tf I: (A,S) J and TIP} . In the
former case, we use the local allocation environment p for evaluating the locality
( and the access list A, while we leave the type 6 unevaluated; it will be evaluated
when the tuple is considered for a matching. If dom{X) C dom{p), we write A{p}
to denote the access list defined by:

Hp}is) = U{A(^) I i e dom{X),p{i) = s}, Vs : 3£ e dom{X) : p{i) = s;

 142

\{p} is not defined whenever dom{X) ^ dom{p). In the latter case, we get a
process closure P{p} that is evaluated by using the laws in Table 9. There,
6{p} is obtained from S by using p to interpret shallow localities of S; it can
be defined inductively on the syntax of 6: all its clauses are structural apart for
(£ H^ TT H->. S'){p} = (.{p] H^ TT H-j. (5'.

r[e]^ = f[el r i ! X1^ = ! X

nP\^ = P{p] Tl\X:5l^ = \X:6
Tll:{\5)l^=p{e):{X{p],5) T{\u:{\,5)l^ = \u:{\{p},6)
nfM, = nf},,nn,
Table 8. Tuple Evaluation Function

The notion of conservative extension of the function used to induce the types
of the nodes of a net is needed to derive the new types of the nodes of the net in
case of dynamic reconfiguration. Indeed, when a set of new nodes is created, the
type specifications written by the programmer have to be taken into account.
In general, the types of the nodes of the net have to be "extended" by adding
the rights of the existing nodes over the new ones and the rights of the new
nodes over the existing ones. Access lists are used to enrich the rights of the
existing nodes, while types are used to determine the rights of the new nodes.
An extension is called conservative whenever the rights over the new nodes are
consistent with the rights over the creating node and the types specified for the
new nodes are compatible with those induced by the extension; the predicate
compat is used to this purpose.

nil{p} = nil
X{p} = X
{out(t)m.p){p} = out{t{p})m{p}.p{p}
{eva.l{Q)m.P){p} = eval(Q)@^{p}.P{p}
{init)m.P){p] = in(t{p})@e{p}.P{p}
{read{t)m.P){p} = rea.d{t{p})m{p}.P{p}
(newloc(u : (xj}).P){p} = newloc(u : (A^)).P{p}
(Pi \ P2){p} = P,{p} \ P2{p}
A(P,I,'S){p} = P[P/X,e/u,e/^{p} iiA{X •.5,u: (x~5},x) ''=^ P

e{p} = e
{i:{\,S)){p}=i{p]:{\{p],5)
!x{p} = ! X
{\u:{\,6)){p) = \u:{X{p],5)
{\X:6){p} = \X:8{p]
{fMp} = f{p},t{p}

Table 9. Closure Laws

 143

As a matter of notation, if S is in canonical form, Sl^i will denote the

capability of locality £, and Sl^ £ the type of i. Both notations can be defined
inductively on the syntax of <5; the more relevant clause, when 6 = £' i-^ TT <-^ S',
is:
f undef iU^e' Si I - I^ ii £ ^ £'
" ^ 1 TT otherwise, [ S' otherwise.
Definition 8. The extension of a type interpretation function © for a set of
sites 5 ' = ( s i , . . . , s„} such that S' n dom{&) = 0 and a site s £ 5, is the type
interpretation function 0 [ ( 0 ( s ) [ s i / s e l f ] ) / s i , . . . , (6'(s)[5„/self])/s„].
The (conservative) extension of a function A that induces the types of the
nodes of a net shall make use of predicate compat, that iteratively on the syntax
of types checks the compatibility of the extension with A and is defined as follows:
def
compat{d, s, 0, A) = Vs' G dom{0) : 6{&]s i^ s' is defined, we have:
A{s){s') C„ S{e}s i, s' and compat{6{9}s i, s', s', 9, A).
Definition 9. Let 5 be a finite set of sites and yl : 5 ^ (5 —^ 77) be a partial
function. The extension A' : (S U {s}) -^ ({S U {s}) -^ 11) of A with respect to
a set of sites {s} such that {s} n 5 = 0, a set of type specifications u : (A, 6)
with cardinality{{u}) = cardinality{{'s}), a type interpretation function O with
dom{&) = 5 U {?} and a site s 6 5, is defined as follows:
- for all si,S2 e 5", we have: A'{si)(s2) = A{si){s2),
- for all Si 6 {s} and sa € 5 U {s}, we have:
A>(, u , ^ - / (^s, [?/S]{0(s)})(52) if A,, ^ 0
lS2JVSij - I ^(S2)(s) otherwise,
- for all si G {s} and 8-2 G S, we have: A'{si){s2) = Ssi['s/u\{0{si)}l^ S2,
whenever the right hand sides of the equalities are defined.
A' is a conservative extension, if it enjoys the following properties:
- for all si G {s} and 52 G S, we have: A'{s2){s) C„ A'(s2){si), A'{s){s2) C.„
A'(si)is2);
- for all si G {s}, we have: A'{s){s) C„ A'{si){si);
- for all si G {s}, we have: compat{Ssi[s/u\,si,0,A').

The structural congruence, =, is the least congruence relation such that:

Nt[[N2=N2\\Nu
{Ni II N2) II 7V3 = iVi II (N2 II N3),
.J
s Pp =_ s4 (Pjnil),
••p ^ - •' --p

s -4 (A IP2) = 5 -4 p, II8 4 P2.

The reduction relation, >—>, is the least relation induced by the laws in
Table 10, where we use the following notations.
Notation 1. In Table 10 the following shorthands are adopted:
— 0Ns, is the type interpretation function associated to Ns' : A, the net before
reduction:
 144

6*^, for Si £ S U {s] U {s}, are the solution of the system of type equations
induced by the conservative extension of A with respect to {s}, u : {X,S), 0
and s, where & is the extension of 0NS, with respect to {s} and s;
Ns[6l./Ss]s-&s is the net Ng where the types 5s of the nodes are replaced
by St},
^{7}' if {«} = {si> • • •, Sn}, is the net Si :://j nil || . . . || s„ ::/,;; nil.

5'=p(£) et = r i t L
Ns II s ::* out(i)@f.P || s' ::*, P ' V ^ iVs || s 4 P \\ s' v."^, ( P ' j out(et))

s = PW
Ns II s 4 eval(Q)@£.P || s' ::f, P ' ^ ^ Ns | h •:' P II s' ^ ( P ' | Q)

s = p{l) 'match{T\t\^,et,s,©Ng,)
Ns II s 4 init)m.P II s' ::f, out(et) ^ - ^ Ns || 3 ::', P [ e t / r [ < ] , ] || s' ::f, nil

s' = p{E) match{Tl t }^,et, s, ON^,)

Ns II s 4 r e a d ( t ) @ l P || s' 4' o^^iet) ^ ^ JVs || s ::* P [ e t / r l i ] J || s' ^ out(et)

{5} n (5 u {«}) = 0
Ns II s ::*: newIoc(S : {X,S)).P y~^ Ns[S:jSs.Ues || s :4f P [ s / ^ || iV
U}

Ns II 5 ::^, P [ P / X J / S , e / J ] ^ iV , , c ; ~ ~ . ~ , ~, rfe/ „
^^.^3 A ( A : d,M : (A,d),a;) = P
N s l h : : ^ A(P,^,S);^yV
N = iVi Ni^-^ Ni N2 = N'
N y-^ N'

Table 10. KLAIM Operational Semantics (see Notation 1)

The pattern-matching predicate used in Table 10 is defined in Table 11. The

matching rules ensure that if a read/in operation looks for processes with type
S then only processes with type S' such that S'{0NS}S '^ S{0Ng}s are accepted.
The rules also ensure that if read/in looks for a site u with type specification
{Xu,Su), then for selecting a site s' with type specification {X,6) it must hold
that:
- the interpretation of (5„[s'/u] at s' is a subtype of both the actual type of s',
6s', and of the interpretation at s' of the type specified for 5', 6{0N^]S'• In
the rules, fl is used to denote the greatest lower bound;
- the capability of a site s" as specified in the input access list, A„(s"), agrees
with that assigned to s" from s' by the coordinator, Ss" 4-^ s'. Moreover, if
the output access list A is not empty, it also agrees with the capability of s"
as specified by A, A(s").
 145

match{v,v, s &Ns) match{P,P,s,&Ns) match(s' : {X,S), s':{\,S), S,0Ns)

match{\x, v, S,0Ns) match{v,! x, s, 0NS )

4>\^P--S' S'{0Ns}s :<S{eNs}s <p[^P:S' S{0Ns}s <S'{0Ns }s

match{ 'X:d,P,s,0Ns) match(P, \ X : 5,i , 0Ns )

Suls'/u]{0Ns}s' :< S^' nS{GNs}s'

Vs" e dom{\u) : iSs" i, s' C „ A„(s")) A (dom(A) / 0 = » X{s") En ^u{s '))
match{\ u : (A„, Su) s' : {X,5},s,0Ns)

du[s' /u]{&Ng }s'^HeNs}s' Vs"G dom{\u) : A(s") C.„ Au(.'")

match{s : {\,5),\u : (A„ ,Su),S,0Ns)

match(fi, f-i s,&Ns} match{t\,t2,s ,0Ns)

match{{fi,ti), (/2,t2), s,0Ns)

T a b l e 1 1 . Matching Rules

References

1. M. Abadi. Secrecy by Typing in Cryptographic Protocols. Theoretical Aspects of

Computer Software (TACS'97), Proceedings (M. Abadi, M. Ito, Eds.), LNCS 1281,
pp.611-638, Springer, 1997.
2. M. Abadi, A.D. Gordon. A calculus for cryptographic protocols: The spi calculus.
Proc. of the ACM Conference on Computer and Communication Security, ACM
Press, 1997.
3. M. Abadi, R. Stata. A Type System for Java Bytecode Verifier. Proc. of the ACM
Symposium on Principles of Programming Languages, ACM Press, 1998.
4. A. Arnold, J. Gosling. The Java Programming Language. Addison Wesley, 1996.
5. L. Bettini, R. De Nicola, G. Ferrari, R. Pugliese. Interactive Mobile Agents in X-
KLAIM. IEEE Seventh International Workshop on Enabling Technologies: Infras-
tructure for Collaborative Enterprises, Proceedings (P. Ciancarini, R. Tolksdorf,
Eds.), IEEE Computer Society Press, 1998.
6. C. Bodei, P. Degano, F. Nielson, H.R. Nielson. Control Flow Analysis for the n-
calculus. Concurrency Theory (CONCUR'98), Proceedings (D. Sangiorgi, R.de Si-
mone, Eds.), LNCS 1466, pp.611-638. Springer, 1998.
7. M. Boreale, D. Sangiorgi. Bisimulation in Naming-Passing Calculi without Match-
ing. Proc. of 13th IEEE Symposium on Logic in Computer Science (LICS '98),
IEEE Computer Society Press, 1998.
8. G. Boudol. Typing the use of resources in a Concurrent Calculus. Advances in
Computing Science (ASIAN'97), Proceedings (R.K. Shyamasundar, K. Ueda, Eds.),
LNCS 1345, pp.239-253, Springer, 1997.
9. L. Cardelli, A. Gordon, Mobile Ambients. Foundations of Software Science and
Computation Structures (FoSSaCS'98), Proceedings (M. Nivat, Ed.), LNCS 1378,
pp.140-155, Springer, 1998.
 146

10. L. Caidelli, A. Gordon, Types for Mobile Ambients. Proc. of the ACM Symposium
on Principles of Program,ming Languages, ACM Press, 1999.
11. N. Carriero, D. Gelernter. Linda in Context. Communications of the ACM,
32(4):444-458, 1989.
12. G. Cugola, C. Ghezzi, G.P. Picco, G. Vigna. Analyzing Mobile Code Languages. In
Mobile Object System,s Towards the Programmable Internet (J. Vitek, C. Tschudin,
Eds.), LNCS 1222, Springer, 1997.
13. R. De Nicola, G. Ferrari, R. Pugliese. Coordinating Mobile Agents via Blackboards
and Access Rights. Coordination Languages and Models (COORDINATION'97),
Proceedings (D. Garlan, D. Le Metayer, Eds.), LNCS 1282, pp. 220-237, Springer,
1997.
14. R. De Nicola, G. Ferrari, R. Pugliese. KLAIM: a Kernel Language for Agents Interac-
tion and Mobility. IEEE Transactions on Software Engineering, Vol.24(5):315-330,
IEEE Computer Society Press, 1998.
15. R. De Nicola, G. Ferrari, R. Pugliese, B. Venneri. Types for Access Control. Avail-
able at h t t p : / / r a p . d s i . u n i f i . i t / p a p e r s . h t m l . To appear in Theoretical Com-
puter Science.
16. D. Gelernter. Generative Communication in Linda. ACM Transactions on Pro-
gramming Languages and Systems, 7(1):80-112, ACM Press, 1985.
17. D. Gelernter, N. Carriero, S. Chandran, et al. Parallel Programming in Linda. Proc.
of the IEEE International Conference on Parallel Programming, pp. 255-263, IEEE
Computer Society Press, 1985.
18. N. Heintz, J.G. Riecke. The SLam calculus; Programming with secrecy and in-
tegrity. Proc. of the ACM Symposium on Principles of Programming Languages,
ACM Press, 1998.
19. M. Hennessy, J. Riely. Resource Access Control in Systems of Mobile Agents. Proc.
Int. Workshop on High-Level Concurrent Languages, vol. 16(3) of Electronic Notes
in Theoretical Computer Science, Elsevier, 1998.
20. N. Kobayashi, B. Pierce, D. Turner. Linearity and the 7r-calculus. Proc. of the
ACM Symposium on Principles of Programming Languages, ACM Press, 1996.
21. R. Milner, J. Parrow, D. Walker. A calculus of mobile processes, (Part I and II).
Information and Computation, 100:1-77, 1992.
22. G. Necula. Proof-carrying code. Proc. of the ACM Symposium on Principles of
Programming Languages, ACM Press, 1997.
23. B. Pierce and D. Sangiorgi. Typing and subtyping for mobile processes. Mathe-
matical Structures in Comp. Science, 6(5):409-454, 1996.
24. J. Riely, M. Hennessy. Trust and Partial Typing in Open Systems of Mobile Agents.
Proc. of the ACM Symposium on Principles of Programming Languages, ACM
Press, 1999.
25. P. Sewell. Global/Local Subtyping and Capabihty Inference for a Distributed
TT-calculus. International Colloquium on Automata, Languages and Programming
(ICALP'98), Proceedings (K.G. Larsen, S. Skyum, G. Winskel, Eds.), LNCS 1443,
Springer, 1998.
26. J. Vitek, G. Castagna. A Calculus of Secure Mobile Computations. Proc. of Work-
shop on Internet Programming Languages, Chicago, 1998.
27. D. Volpano, G. Smith. A typed-based approach to program security. Theory
and Practice of Software Development (TAPSOFT'97), Proceeding (M. Bidoit,
M. Dauchet, Eds.), LNCS 1214, pp.607-621, Springer, 1997.
28. D. Volpano, G. Smith. Secure Information Flow in a Multi-threaded Imperative
Language. Proc. of the ACM Symposium on Principles of Programming Languages,
ACM Press, 1998.
 Security Properties of Typed Applets

Xavier Leroy and Prangois Rouaix

Abstract. This paper formcdizes the folklore result that strongly-typed

applets are more secure than untyped ones. We formulate and prove
several security properties that all well-typed applets possess, and iden-
tify sufficient conditions for the applet execution environment to be safe,
such as procedural encapsulation, type abstraction, and systematic type-
based placement of run-time checks. These results are a first step towards
formal techniques for developing and validating safe execution environ-
ments for applets.

1 Introduction
What, exactly, makes strongly-typed applets more secure than untyped ones?
Most frameworks proposed so far for safe local execution of foreign code rely
on strong typing, either statically checked at the client side [19,44], statically
checked at the server side and cryptographically signed [35], or dynamically
checked by the client [8]. However, the main property guaranteed by strong typ-
ing is type soundness: "well-typed programs do not go wrong", e.g. do not apply
an integer as if it were a function. While violations of type soundness constitute
real security threats (casting a well-chosen string to a function or object type
allows arbitrary code to be executed), there are many more security concerns,
such as integrity of the running site (an applet should not delete or modify ar-
bitrary files) and confidentiality of user's data (an applet should not divulge
personal information over the network). The corresponding security violations
do not generally invalidate type soundness in the conventional sense.
If we examine the various security problems identified for Java applets [12],
some of them do cause a violation of Java type soundness [21]; others corre-
spond to malicious, but well-typed, uses of improperly protected functions from
the applet's execution environment [7]. Another typical example is the ActiveX
applet described in [10] that does a Trojan attack on the Quicken home-banking
software: money gets transferred from the user's bank account to some offshore
account, all in a perfectly type-safe way.
On these examples, it is intuitively obvious that security properties must
be enforced by the applet's execution environment. It is the environment that
eventually decides which computer resources the applet can access. This is the
essence of the so-called "sandbox model". Strong typing comes in the picture
only to guarantee that this environment is used in accordance with its publicized
interface. For instance, typing prevents an applet from jumping in the middle
of the code for an environment function, or scanning the whole memory space
of the browser, which would allow the applet to abuse or bypass entirely the
execution environment.
 148

The purpose of this paper is to give a formal foundation to the intuition

above. We formulate and prove several security properties that all well-typed
applets possess. Along the way, we identify sufficient conditions for the execu-
tion environment to be safe, such as procedural encapsulation, type abstraction,
and systematic type-based placement of run-time checks. These results are a
first step towards formal techniques for developing and validating safe execution
environments for applets.
The remainder of this paper is organized as follows. Section 2 introduces a
simple language for applets and their environment, and formalizes the security
policy that they must obey. Section 3 proves a security property based on the
notion of lexical scoping, then extends it to take procedural abstraction into
account. In section 4, we equip our language with a simple type system, which
is used in section 5 to prove three type-based security properties, two relying
on run-time checks and the other on a combination of run-time checks and type
abstraction. After a brief parallel with object-oriented languages in section 6,
section 7 outlines the security architecture for applets in the MMM Web browser
and discusses the relevance of our results to this practical example. Related work
is discussed in section 8, followed by concluding remarks in section 9.

2 The language and its security policy

2.1 The language
The language we consider in this paper is a simple lambda-calculus with base
types (integers, strings, . . . ) , pairs as the only data structure, and references in
the style of ML. The syntax of terms, with typical element a, is as follows:
Terms: a ::= x identifiers
Ib constant of base type (integer, . . . )
I Aa;.a function abstraction
I 01(02) function application
I (01,02) pair construction
I f st(o) I snd(o) pair projections
I ref (o) reference creation
I !o dereferencing
I Oi := 02 reference assignment
References are included in the language from the beginning not only to account
for imperative programming (all kinds of assignment on variables and mutable
data structures such as arrays can easily be modeled with references), but also
to provide easily observable criteria on which we base the security policy.

2.2 The security policy

The security policy we apply to applets is based on the notion of sensitive store
locations: locations of references that an applet must not modify during its ex-
ecution, or more generally references that can be modified during the applet
 149

execution, but whose successive contents must always satisfy some invariant, i.e.
remain within a given set of permitted values.
The first motivation for this policy is to formalize the intuitive idea that an
applet must not trash the memory of the computer executing it. In particular,
the internal state of the browser, the operating system, and other applications
running on the machine must not be adversely affected by the applet.
This security pohcy can also be stretched to account for input/output behav-
ior, notably accesses to files and simple cases of network connections. A low-level,
hardware-oriented view of I/O is to consider hardware devices such as the disk
controller and network interface as special locations in the store; I/O is then
controlled by restricting what can be written to these locations. For a higher-
level view, each file or network connection can be viewed as a reference, which
can then be controlled independently of others. Here, the file system, the name
service and the routing tables become dictionary-like data structures mapping
file names, host names, and network addresses to the references representing files
and connections.
By concentrating on writes to sensitive locations, we focus on integrity prop-
erties of the system running the applet. It is also possible to control reads from
sensitive locations, thus estabhshing simple privacy properties. We will not do
it in this paper for the sake of simplicity, but the results of section 3 also extend
to controlled reads. More advanced privacy properties, as provided for instance
by information flow models, are well beyond our approach, however.

2.3 The instrumented semantics

To enforce the security policy, we give a semantics to our language that monitors
reference assignments, and reports run-time errors in the case of illegal writes.
We use a standard big-step operational semantics in the style of [32,39,25].
Source terms are mapped to values, which are terms with the following syntax:
Values: •.= b values of base types
I Aa;.o[e] function closures
pairs of values
store locations
Results: := v/s normal termination
I err write violation detected
Environments: e : = [xi <(- vi

Stores: s := [£i ^ VI .

The evaluation relation, defined by the inference rules in Fig. 1, is written

(/J, e,s h o —>^ r, meaning that in evaluation environment e, initial store s, and
store control (p, the source term a evaluates to the result r, which is either err
if an illegal write was detected or a pair v/s' of a value v for the source term
and a modified store s'.
 150

Normal rules:

(/9,e, s f-x ^ e(x)/s (1) if,e,s\-b—^b/s (2)

(p,e,s h Xx.a—^ \x.a[e]/s (3)
ip,e,s \- ai -^ Xx.a'[e']/si <p, e, si h aa —>• U2/S2 ip,e'{x *-V2},S2 \-a —¥ r
(4)
<p,e,s h ai(a2) —> r
ip,e,s h ai -^ vijsi tp, e, si h a2 —> 1^2/32
(5)
ip,e,s\- (01,02) -> {Vl,V2)/S2

¥?, e, s h o —)• (ui, V2)/s' ij5, e, s I- o —> (ui, V2)/s'

(6) (7)
V, e, s I- f s t ( a ) -4 vi/s' <p,e,s h snd(a) —> V2Is'
ip,e,s\- a ^ v/s' ^ ^ Dom(s') UDom(v?) ip,e,s\- a -¥ Ij s'
(8) (9)
<p,e,s h r e f (o) -> e.ls{i<r-v} <p,e,s h !a-> s'{t)ls'
</j, e, s h oi ->• ijsi tp, e, si h 02 ->• V2/S2 ^ ^ D o m ( v ) or V2 6 ¥'(^)
(10)
(^, e,s t- (oi := 02) —> O/s2{i i- vz}

tp,e,sh ai —> f/si (/J, e, si h 02 —> 1^2/32 ^ € Dom(i/3) and V2 ^ v'(^)

(11)
ip,e,s\- (oi ;= 02) —>• err

Error propagation rules:

(p,e,s \-ai —> en ip,e,s\-ai -^ vi/si (^, e, Si h 02 - + e r r

(13)
V.e,s\- a 1(02) ->• err (/), e, s I- oi (02) —>• err

V , e, s 1- oi —> err yi, e, s I- Oi —> ?;i/si ip, e, si I- 02 —>• err

-(14) (15)
V^e, s \- (01 ,02) —)• err ifi,e,s h- (oi, 02) —> err

•^ e, s 1- a ->err 1^, e,s h o -4 err

(17)
^,e, s h fst(a) —^ err ip,e,s \- snd(a) —> err
ip,e.,s \- a —> err </3,e, s h a -+ err
(18) • (19)
ifi,e,s h ref (a) —)• err v', e, s h !o —> err
(p, e, s h oi —^ err <p, e,s h oi -> n i / s i yj, e, si I-02 —> err
(20) ; ; (21)
(/?, e, s h (oi := 02) —>• err (p,e,s\- (oi := 02) -> err

Fig. 1. Evaluation rules

 151

The only unusual ingredient in this semantics is the <^ component, which
maps store locations to sets of values: if (p(£) is defined, values written to the
location i must belong to the set '^{i), otherwise a run-time error err is gener-
ated; if i{>{i) is undefined, any value can be stored at i. (See rules 10 and 11.)
For instance, taking ^piji) = 0 prevents any assignment to i.
The rules for propagating the e r r result and aborting execution (rules 12-
21) are the same rules as for propagating run-time type errors (wrong) in [39];
the only difference is that we have no rules to detect run-time type errors, thus
making no difference between run-time type violations and non-terminating pro-
grams (no derivations exist in both cases): the standard type soundness theorems
show that type violations cannot occur at run-time in well-typed source terms.
An unusual aspect of our formalism is that the store control <f must be
given at the start of the execution. The reason is that, with big-step operational
semantics, it does not suffice to perform a regular evaluation e,s h a —> w/s'
and observe the differences between s and s' to detect illegal writes. For one
thing, we would not observe temporary assignments, where a malicious applet
writes illegal values to a sensitive location, then restores the original values before
terminating. Also, we could not say anything about non-terminating terms: the
applet could perform illegal writes, then enter an infinite loop to avoid detection.
By providing the store control i^ in advance, we ensure that the first write error
will be detected immediately and reported as the e r r result.
Unfortunately, this provides no way to control stores to locations created
during the evaluation (rule 8 chooses these locations outside of Dom((p), meaning
that writes to these locations will be free): only preexisting locations can be
sensitive. (This can be viewed as an inadequacy of big-step semantics, and a
small-step, reduction-based semantics would fare better here. However, several
semantic features that play an important role in our study are easier to express in
a big-step semantics than in a reduction semantics: the clean separation between
browser-supplied environment and applet-supplied source term, and the ability
to interpret abstract type names by arbitrary sets of values.)

3 Reachability-based security

3.1 Simple reachability

The first security property for our calculus formalizes the idea that an applet can
only write to locations that are reachable from the initial environment in which
it executes, or that are created during the applet's execution. For instance, if the
references representing files are not reachable from the execution environment
given to applets, then no applet can write to a file.
Reachability, here, is to be understood in the garbage collection sense: a
location is reachable if there exists a path in the memory graph from the initial
environment to the location, following one or several pointers. More formally, we
define the set RL(v, s) of locations reachable from a value w in a store s by the
 152

following equations:

RL{b, s) = 0
RL{\x.a[e],s) = RL{e,s)
RL{{vi,V2),s) = RL{vi,s)URL{v2,s)
RL{£,s) = {e}URL{s{^),s)
RL{e,s)=^ [j RL{e{x),s)
xeDom(e)

The definition above is not well-founded by induction on v, since in the fourth

case the value s{£) is arbitrarily lajge and may contain £ again. It should be
viewed as a fixpoint equation RL = F{RL), where F is an increasing operator;
RL is, then, the smallest fixpoint of that operator.
If p is a set of sensitive locations, we define Prot{p) as the store control that
maps locations ^ G p to 0, thus disallowing all writes on £, and is undefined on
locations £ ^ p- We can now formulate the first security property:
Security property 1 Let p be a set of sensitive locations. Ifpr\RL{e,s) = 0,
then for all applets a, we have Prot(p),e, s\- a -/^ e r r .
In other words, if none of the locations in p is reachable from the initial
environment e and store s, then no applet a can trigger an error by writing to
a location in p: the applet will either terminate normally or loop. The proof of
this property is a simple inductive argument on the evaluation derivation, using
the following lemma as the induction hypothesis:
Lemma 1. If pr\RL{e, s) = 0 and Prot{p), e,s h a —^ r, then r ^ e r r . Instead,
r = v/s' for some v and s'. Moreover, p fl RL{v, s') = 0, and for all values w
such that p n RL{w, s) = 0, we have p fi RL{w, s') = 0 .
It is important to remark that Property 1 holds in practice only if the evalua-
tion of a on an actual processor adheres to the rules given in Fig. 1. For instance,
the rules do not allow the evaluation of b := a where b is an integer constant,
but an untyped or weakly-typed implementation might execute b := a without
crashing nor reporting an error in the case where 6 is a valid memory address;
this would allow the applet to write to arbitrary locations. We rely on a being
evaluated by a type-safe implementation to ensure that this cannot happen: ei-
ther the evaluator must perform run-time type-checking, or a must be well-typed
in some sound static type system (such as the one presented below in section 4).
Property 1, though simple and already well-known in the field of garbage
collection [26], establishes a number of properties without which no security is
possible at all. First, our language has safe pointers: locations cannot be forged by
casting well-chosen integers. Second, automatic memory management (garbage
collection) is feasible and does not weaken security: a location that becomes un-
reachable remains unreachable; moreover, newly allocated locations are always
initialized; therefore, unreachable locations can be reused safely. Third, the lan-
guage enforces lexical scoping: the execution of the applet depends only on the
 153

environment in which it proceeds; the applet does not have access to the full
execution environment of the browser — as would be the case in a dynamically-
scoped language, such as Emacs Lisp, or a language with special constructs to
access the environment of the caller, such as Tel.

3.2 Reachability and procedural abstraction

In defining reachable locations, we have treated closures like tuples (as garbage
collectors do): the locations reachable from Ax.a[e] are those reachable from e{y)
for some y 6 Dom(e). There is, however, a big difference between closures and
tuples. Tuples are passive data structures: any piece of code that has access to
the tuple can then obtain pointers to the components of the tuple. Closures are
active data structures: only the code part of the closure can access directly the
data part of the closure (the values of the free variables); other code fragments
can only apply the closure, but not access the data part directly. In other words,
the code part of a closure mediates access to the data part. This property is
often referred to as procedural abstraction [34].
For instance, consider the following function, similar to many Unix sys-
tem calls, where uid is a reference holding the identity of the caller (applet
or browser):

Ax. if !uid = browser

then do something with high privileges
e l s e do something with low privileges

Assume this function is part of the applet environment, but not the reference
uid itself. Then, there is no way that the applet can modify the location of uid,
even though that location is reachable from the environment.
A less obvious example, where the reference uid is not trivially read-only, is
the following function in the style of the Unix s e t u i d system call:
Anewid. if !uid = browser
then uid := newid
e l s e raise an error
Assuming uid is not initially browser, an applet cannot change uid by calling
this function.
Procedural abstraction can be viewed as the foundation for access control
lists and similar programming techniques, which systematically encapsulate re-
sources inside functions that check the identity and credentials of the caller
before granting access to the requested resources. For instance, a file opening
function contains the whole data structure representing the file system in its
closure, but grants access only to files with suitable permissions. Thus, while all
files are reachable from the closure of the open function, only those that have
suitable permissions can be modified by the caller.
To formalize these ideas, we set out to define the set of locations ML{v, s)
that are actually modifiable (not merely reachable) from a value v and a store s,
 154

and show that if a location i is not in ML{e,s), then any applet evaluated in
the environment e does not write to i. This result is stronger than Property 1
because the location £ that is not modifiable from e in s can still be reachable
(via closures) from e in s.
For passive data structures (locations, tuples), modifiability coincides with
reachabihty: a location is modifiable from v/s if a sequence of f s t , snd, and ! op-
erations applied to u in s evaluates to that location. The difficult case is defining
modifiable locations for a function closure. The idea is to consider all possible
applications of the closure to an argument: a location I is considered modifiable
from the closure only if one of those applications writes to the location, or causes
the location to become modifiable otherwise.
More precisely, let Capi be the execution environment given to applets, and
let c = Aa;.a[e] be one of the closures contained in Capi- A location £ is modifiable
from c in store s if there exists a value v such that the following conditions hold:
Condition 1. The application of the closure to v causes £ to be modified, i.e.
{£ !-)• 0}, e{x •(- u}, s h o -> e r r .
Example: Let e be the environment [r i- i]. Then, i is reachable from {Xx. r :=
a; + l)[e] in any store, since any application of the closure causes i. to be assigned.

Condition 2. Alternatively, the apphcation of c to t; does not modify i, but

returns a result value and a new store from which £ is modifiable.
Example: With e as in the previous example, £ is reachable from (Aa;. (r, l))[e],
since any application of that closure returns a pair with £ as first component.
An applet can thus write to £ by applying the closure, extracting £ from the
returned result, and assigning £ directly.

Condition 3. Alternatively, the application of c to w modifies a reference acces-

sible to the applet in such a way that £ becomes modifiable from that reference.
Example: Consider c = {\p. p := r)[e], where e = [r ^ ^] as usual. Applying c
to a location £', we obtain a modified store s' such that s'{£') = £, i.e. £ is
modifiable from £' in s' and can be modified by the applet.

Condition 4. Alternatively, the application of c to u assigns references internal

to the browser functions in such a way that £ becomes modifiable from the
environment Capi in the store s' at the end of the application.
Example: Consider the following environment eapi-

eapiif) = (Aa;. n := \n + l)[n <- £„]

Capiig) — {Xx. if !n > 1 then r := 0)[n -f- £„;r f- £]

In a store s such that s(£„) = 0, the location £ is not modifiable from Sapiig)-
However, £ is modifiable from e„pi(/) in s, since one application of that closure
returns a store s' such that s'(£„) = 1, and in that store s', t \s modifiable from
Sapiig)'- ^•iiy application of enpi{g) with initial store s' writes to £.
 155

Condition 5. In conditions 1-4 above, it must be the case that the location t
found to be modifiable from the closure c is not actually modifiable from the
argument v passed to the closure. Otherwise, we would not know whether the
location really "comes from" the closure c, or is merely modified by the applet-
provided argument v.
Example: Consider the higher-order function c = (A/. /(O))[0]. If we apply c to
(An. r := Vb)\r <- £], we observe a write to location I. However, i should not be
considered as modifiable from c, since it is also modifiable from the argument
given to c.
As should now be apparent from the conditions 1-5 above, the notion of mod-
ifiability raises serious problems, both practical and technical. On the practical
side, the set of modifiable locations ML{v, s) is not computable from v and s: in
the closure case, we must consider infinitely many possible arguments. Thus, a
full mathematical proof is needed to determine ML{v,s).
Moreover, modifiable locations cannot be determined locally. As condition 4
shows, the modifiable locations of a closure depend on the modifiable locations
of all functions from the applet environment Capi- Thus, if we manage to deter-
mine ML{eapi,s), then add one single function to the applet environment, we
must not only determine the modifiable locations from the new function, but
also reconsider all other functions in the environment to see whether their mod-
ifiable locations have changed. This is clearly impractical. Hence, the notion of
modifiability is not effective and is interesting only from a semantic viewpoint
and as a guide to derive decidable security criteria in the sequel.
On the technical side, the conditions 1-5 above do not lead to a well-founded
definition of the sets of modifiable locations ML{v, s). The problem is condition 5
(the requirement that the location must not be modifiable from the argument
given to the closure): viewing conditions 1-4 as a fixpoint equation for some
operator, that operator is not increasing because of the negation in condition 5.
In appendix B, we tackle this problem and show that non-modifiable locations
are indeed never modified in the particular case where the applet's environment
Bapi is well-typed and its type Eapi does not contain any ref types, so that no
references are exchanged directly between the applet and its environment. In
the remainder of this paper, we abandon the notion of modifiability in its full
generality, and develop more effective techniques to restrict writes to reachable
locations, relying on type-based instrumentation of the browser code.

4 The type system

We now equip our language with a simple type system. The type system is based
on simply-typed A-calculus, with the addition of named, user-defined types. De-
spite its simplicity, this type system does not restrict drastically the expressive-
ness of our language. In particular, recursive functions can still be defined using
references [39]. The type algebra is:
 156

Types: r v.— i base type (int, s t r i n g , etc.)

It named type
I Ti —> r2 function type
I Tj X r2 product type
I T ref reference type
Conversely, we enrich the syntax of terms with two new constructs: exphcit
coercions to and from named types, and run-time vahdation of values of named
types.
Terms: a::— ... (as before)
I r{a) coercion to type r
I OKtia) run-time validation at type t
The typing rules for the calculus are shown in Fig. 3, and the extra evaluation
rules for the new constructs in Fig. 2.

(p, e, s I- a -^ r
(22)
ifi,e,s\- T{a) —> r
<fi,e,s\~ a ^ v/s t 6 Dom(PV) implies v 6 PV{t)
(23)
(p,e,s h OKt{a) -> v/s
f,e,s\-a-^ err
(24)
ip,e,s\- OKt(a) —>• err

Fig. 2. Extra evaluation rules for coercions and run-time checks

4.1 Named types

The overall approach followed in section 5 is to identify groups of references

having the same type, and apply a given security policy to all of them. However,
types from the simply-typed A-calculus are too coarse for this purpose: references
of type s t r i n g ref can hold many different kinds of data, such as messages,
filenames, and cryptographic keys; clearly, different security restrictions must be
applied to these different kinds of strings.
To this end, we introduce named types t, defined by a mapping TD for type
names to type expressions, stating that the type t is interconvertible with its im-
plementation type TD(t). For instance, we could introduce a type f ileneune,
defined to be equal to s t r i n g . An expression e of type s t r i n g cannot be
used implicitly with type filename; an explicit injection into filename, written
f ilename(e), is required. Conversely, accessing the string underlying an expres-
sion e' of type filename is achieved by s t r i n g ( e ' ) . (See rules 35 and 36.) In
 157

Eh X : E{x) (25) Ehb: Typeof{b) (26)

E{x <-T'}\- a-.T Ehai-.r'-^T E \-a-z : T'
(27) (28)
E h Xx.a : T' ^ T EV- 01(02) : r
E\- ai : Ti E h a2 : T2 E \- a : TI x T2 E \- a : n x T2
(29) (30) (31)
E \- ( a i , 02) •• Ti X T2 E\- fst(a) : n E \- snd(o) : T2
Eb a: T E\- a: T ref
(32) • (33)
£; h ref(a) : r ref E h-\a : T
E \- ai : T ref E \- a2 : r
(34)
E h {a\ := az) : u n i t
E\- a-.T T= TD{t) Eh a:t T = TD{t)
(35) (36)
E V- t{a) : t Eh T{a) : T
Eh a:t
(37)
E h OKt(a) • t

Fig. 3. Typing rules

this respect, our named types behave very much like the i s new type definition
in Ada, and unhke type abbreviations in ML. Making the coercions exphcit fa-
cihtates the definition of the program transformations in section 5, ensuring in
particular that each term has a unique type.
The mapping TD of type definitions is essentially global: type definitions
local to an expression are not supported. Still, it is possible to type-check some
terms against a set of type definitions TD' that is a strict subset of TD, thus
rendering the named types not defined in TD' abstract in that term. We will use
this facility in section 5.2 and 5.3 to make named types abstract in the applet.

4.2 Run-time validation of values

The other unusual feature of our type system is the family of operators OKt (one
for each named type t) used to perform run-time validation of their argument.
For each named type t, we assume given a set PV{t) of permitted values for
type t. (We actually allow PV{t) to be undefined for some types t, which we
take to mean that all values of type t are valid.) The expression OKt{e) checks
whether the value of e is in PV{t); if yes, it returns the value unchanged; if
no, it aborts the execution of the applet and reports an error. In the evaluation
rules, the "yes" case corresponds to rule 23; there is no rule for the "no" case,
meaning that no evaluation derivation exists if an OKt test fails. In effect, we
do not distinguish between failure of OKt and non-termination. At any rate, we
 158

must not return the e r r result when OKt fails: no write violation has occurred
yet.
By varying PV^t), we can control precisely the values of type t that will pass
run-time validation. For instance, Py(filename) could consist of all strings
referencing files under the applet's temporary directory /tmp/applet.x, a new
directory that is created empty at the beginning of the applet's execution. Com-
bined with the techniques described in section 5, this would ensure that only
files in this temporary directory can be accessed by the applet. Similarly, the
set P y (widget) could consist of all GUI widget descriptors referring to widgets
that axe children of the applet's top widget, thus preventing the applet from
interacting with widgets belonging to the browser. Other examples of run-time
validation include checking cryptographic signatures on the applet code itself or
on sensitive data presented by the applet.
In practice, validation OKt{e) involves not only the value of its argument e,
but also external information such as the identity of the principal, extra capa-
bility arguments passed to the validation functions, and possibly user replies
to dialog boxes. A typical example is the Java SecurityManager class, which
determines the identity of the principal by inspection of the call stack [43]. For
simphcity, we still write OKt as a function of the value of its argument.
The evaluation rule for OKt assumes of course that membership in PV{t)
is decidable. This raises obvious difficulties if t stands for a function type, at
least if the domain type is infinite. Difficulties for defining PV{t) also arise if t
is a reference type: checking the current contents of the references offers no
guarantees with respect to future modifications; checking the locations of the
references against a fixed set of locations is very restrictive. For those reasons,
we restrict ourselves to types t that are defined as algebraic datatypes: type
expressions obtained by combining base types with datatype constructors such
as l i s t or tuples, but not with ref nor the function arrow.

4.3 Type soundness

To relate the typing rules to the dynamic semantics, we define a semantic typing
relation S \=: v : T saying whether the value v is a semantically correct value for
type r. The S component is a store typing, associating types to store locations.
We simultaneously extend the |= relation to stores and store typings (|:= s : S),
and to evaluation environments and typing environments (5 [= e : E). The
definition of |= is shown below, and is completely standard [39,23].

- S 1= 6 : t if Typeofib) = i
- S[=v:tiiS^v: TD{t)
- S \= Xx.a[e\ : Ti —> T2 if there exists a typing environment E such that
S \=^ e : E and E h Xx.a : ri —^ r2
- 5 1= (wi, v-i) : Ti X r2 if S 1= vi : Ti and S \= v-z : T2
- 5 1= £ : T ref if T = S{t)
- S\=e:E\i Dom(e) = Dom(£) and for all x € Dom(e), S [= e{x) : E{x)
- 1= s : 5 if Dom(s) == Dom(5) and for all I e Dom(s), 5 f= s{(.) : 5(£).
 159

Using the semantic typing relations defined above, we then have the familiar
strong soundness property below for the type system. We say that a store typing
S" extends another store typing S if Dom(5") 3 Dom(5), and for all £ E Dom(5),
we have S'{C) = S{f.). Remark that semantic typing is stable under store exten-
sion; ii S \= V • T and S' extends S, we also have S' ]= v : T.

Lemma 2 (Type soundness). Assume E \- a : T and \= s : S and S \== e : E.

If (p,e,s h a -^ v/s', then there exists a store typing S' extending S such that
S' \= V : T and |= s' : S'.

Proof. The proof is a simple inductive argument on the evaluation derivation;

see [23, Prop. 3.6] for details.

5 Type-based security properties

The type-based security properties we develop in this section are based on a
common idea: assuming all sensitive references have types of the form t ref, we
instrument the functions composing the execution environment by inserting OKt
run-time checks at certain program points, in order to prevent illegal writes to
references of type t ref. The applets themselves are not instrumented, of course,
since their source code is generally unavailable. All we know about the applet is
that it is well typed in a given typing environment and set of type definitions.
It is the combination of this well-typing with the instrumented environment
functions that guarantees security.
To illustrate the three instrumentation schemes proposed below, we use pic-
tures in the style of Fig. 4 showing the flow of values between the applet and its
execution environment. We focus on values of a named type t whose implemen-
tation type is r = TD{t). The goal of our instrumentation schemes is to make
sure that only checked values of type t (solid arrows in the pictures) can reach
an assignment on a i ref performed by the execution environment.

5.1 Instrumentation of writes and procedural abstraction

The first transformation we consider inserts an OKt check before any write to
a reference of type t ref with t G Dom(Py). This way, we are certain that if a
sensitive reference has type t ref, the environment functions will always store in
it values that belong to PV{t). See Fig. 5 for a graphical illustration. Formally,
we define the instrumentation scheme IW, operating on terms a^ annotated
with their type r, as follows:

IW{{a^ " ' := 6*)™") = IW{a^ " ' ) := OKt{IW{b^)) if t e Dom(FF)

On other kinds of terms, IW is a simple morphism, e.g. IW{{Xx.a^Y^'^) =

Xx.IWiaT)., etc.
It is easy to show that write errors cannot happen inside instrumented terms.
Given the permitted values PV and a store typing S, we define Prot{PV,S) as
 160

applet environment

value of static type r (e.g. s t r i n g )

unchecked value of static type t (e.g. unchecked filename)
checked value of static type t (e.g. checked filename)

Fig. 4. Flow of values between the applet and its environment

applet environment

Fig. 5. Instrumentation of writes and procedural abstraction

»{i(?)--»{o^

-(my*-~i^{^-
applet environment

Fig. 6. Instrumentation of coercions ajid type abstraction

applet environment

Fig. 7. Instrumenting coercions without type abstraction

 161

the store control that restricts references of type t ref, t £ Dom.{PV) to values in
PV{t), and allows arbitrary writes to other references: Prot{PV,S){l) = PV{t)
if 5(£) = i and t e Dom(Py), and Prot,{PV, S){e) is undefined otherwise.

Lemma 3. Assume E h (a^ "' := b'^) : u n i t and S \= e : E and [= s : S. Define

if = Prot{PV,S). If^,e,s h IWia"- "') -/^ e r r and ip,e,sh- IWi¥) 7^ err,
then ip,e,s\- IW{a'' " ' := 6^") A e r r .

Proof. If r = i for some named type t 6 Dom(Py), then by definition of the

instrumentation scheme, the right-hand side of the assignment is of the form
OKt{b') for some 6', which can only evaluate to an element of PV{t) by rule 23.
Hence, the assignment is valid with respect to Prot{PV,S). If r is not a t type
or is outside Dom(Py), by Lemma 2, IW{a) evaluates to a location i which
does not have a type of the form t ref, t € Dom(PF), and therefore such that
Prot{PV, S){i) is undefined; hence, no write error occurs either.

Lemma 3 only provides half of the security property: it shows that writes
in instrumented code are safe, but only the execution environment contains
instrumented code; the applet code is not instrumented and could therefore
perform illegal writes to sensitive locations, if it could access those locations.
In other terms, we must make sure that all sensitive locations are encapsulated
inside functions, as in section 3.2. To this end, we will restrict the type Eapi of the
applet's execution environment Sapi to ensure that sensitive references cannot
"leak" into the applet, and be assigned illegal values there. There are several
ways by which a sensitive reference of type t ref could leak into an applet:

— The reference is exported directly in the environment, e.g. Eapi{x) = t ref

or Eapi{x) — i n t x (i ref).
— The reference is returned by one of the functions of the environment, e.g.
Eapiif) = int-^t ref.
— The environment contains a higher-order function such as Eapi{h) =
{t ref -> i n t ) —^ i n t . The applet could get access to a sensitive reference
if h passes one to its functional argument, which can be provided by the
applet.
— The environment contains a function taking as argument an applet-provided
reference to a f ref, e.g. Eapiif) = t ref ref -^ imit. The environment
function / could then store a sensitive location into that t ref ref, from
which the applet can recover the sensitive location later.

We rule out all these cases by simply requiring that no type t ref occurs (at any
depth) in Eapi- This leads to the following security property:

Security property 2 Assume S \= eapi • Eapi and \= s : S. Further assume

that Eapi contains no occurrence oft ref for any t, and that all function closures
in e and s have been instrumented with the IW scheme (that is, e and s are
obtained by evaluating source terms instrumented with IW). Then, for every
applet a well-typed in E, we have Prot{PV, S), s,e \- a -/^ err.
 162

Proof. We consider all assignments to references that occur in the evaluation

derivation for a. By Lemma 3, assignments performed by environment functions
cannot cause a write error, since the right-hand side has been instrumented
by IW. For assignments performed by the applet, we use a containment lemma
developed in appendix A to show that the reference being assigned cannot belong
to T)ora{Prot{PV, S)). More precisely, we apply the results of appendix A with T
being the set of all type expressions where t ref occurs, Lapp being the set of
locations with types in T allocated by the applet, and Lgm, being the set of
locations with types in T allocated by environment functions or initially present
in &api- The containment lemma (Lemma 7) then shows that the location being
assigned does not belong to Lenv Moreover, by construction of Prot and Le„„,
we have Dom(Proi(Py, 5)) C Lenw Thus, assignments performed by the applet
do not cause write violations either. Hence, the applet cannot evaluate to err.

The requirement that no t ref occurs in Eapi is clearly too strong: nothing
wrong could happen if, for instance, one of the environment functions has type
t ref -^ u n i t (the t ref argument is provided by the applet). We conjecture that
it suffices to require that no type t ref occurs in Eapp at a positive occurrence
or under a ref constructor. However, our proof of Property 2, and in particular
the crucial containment lemma, does not extend to this weaker hypothesis.

5.2 Instrumentation of coercions and type abstraction

Instead of putting run-time checks on writes to references of type t ref, we
can ensure that all values of type t £ Dom(Py) that flow through the applet's
execution environment always belong to P'V{t). This way, values stored in a
t ref will automatically satisfy PV{f) as well. This is achieved by adding checks
to all creations of values of type t € T>om.{PV) in the execution environment, i.e.
to coercions of the form t{a), following the instrumentation scheme IC below:

IC{t{a)) = OKt{t{IC{a))) if t £ Dom{PV)

Of course, this is not enough: the applet could forge unchecked values of type t,
by direct coercion from t's implementation type, and pass them to environment
functions. Hence, we also need to make the types t e Dom(Py) abstract in the
applet, by type-checking it with a set of type definitions TD' obtained from
TD by removing the definitions of the types t £ Dom{PV). Then, for any
t € Dom(PV), the only values of type t that can be manipulated by the applet
have been created and checked by the environment. This is depicted in Fig. 6.
To capture the run-time behavior of instrumented terms, we introduce a
variant of the semantic typing predicate, written PV, S \= v : T, which is similar
to the predicate S \= v : T from section 4.3, with the difference that a value v
belongs to a named type t only ii v £ PV{t) in addition to v belonging to the
definition TD{t) of t:
- PV,S\=b: i if Typeof{h) = i
- PV,S\=v:t\iPV,S\=v: TD{t) and t € V)om{PV) implies v G PV{t)
 163

- PV, S \= Xx.a[e] : TI -> rs if there exists a typing environment E such that

PV, S \= e : E and E h- Xx.a : n -> rj
- P F , 5 1= (vi.Wi) : ri X r2 if PV,S\= Vi : n and PV,S\= V2 : T2
- P y , 5 1= € : r ref if r = S{1)
- PV,S\=e:Eii Dom(e) = Dom(£;) and for all x e Dom(e), PV, S \= e{x) :
E{x)
- P y 1= s : 5 if Dom(s) = Dom(5) and for all (. € Dom(s), PV,S |= s{(.) :
5(£).
We then have the following characterization of the the behavior of terms instru-
mented with IC:

Lemma 4. Assume E \- a : T and PV,S |= e : £ and PV \= s : S. Further

assume that all closures contained in e and s have function bodies instrumented
with IC. If Prot{PV,S),e,s \- IC{a) —> r, then r ^ e r r ; instead, r is of the
form v/s', and there exists a store txjping S' extending S such that PV, 5" |= w : r
andPV^s' : S'.

Compared with Lemma 2, we now use the more restrictive interpretation of

named types PV, and require that all terms occurring in the evaluation deriva-
tion have been instrumented with IC. The proof is essentially identical to that
of Lemma 2.
Notice that an applet a well-typed within the restricted set TD' of type
definitions contains no coercions t{a) for any t 6 Dom(Py), hence is equal
to IC{a). Thus, Lemma 4 applies not only to the terms from the execution
environment, but also to the applet itself. In a sense, this lemma can be viewed
as a parametricity result, in that it shows soundness for arbitrary interpretations
PV of the types left abstract in TD'. Fiom this remark, we immediately obtain
the following security property:

Security property 3 Let e be the execution environment for applets, and s the
initial store. Assume that all function closures in e and s have been instrumented
with the IC scheme (that is, e and s are obtained by evaluating source terms
instrumented with IC). Assume PV \= s : S and PV, S [^ e : E. Then, for
every applet a well-typed in E and in the restricted set TD' of type definitions,
we have Prot{PV, S), s,e\- a -/^ e r r .

Property 3 can be viewed as a formal justification for capability-based sys-

tems: by making the type of capabilities abstract to the applets, run-time security
checks are necessary only at points where new capabilities are constructed and
returned to the applet; capabilities presented by the applet can then be trusted
without further checks.
Unlike Property 2, Property 3 does not require that types t ref do not occur
in the typing environment E. Indeed, once t is made abstract, it is perfectly safe
to make references of type t ref accessible to the applet: the applet can then
write to them, but only write safe values of type t. Hence, sensitive references no
longer need to be systematically wrapped inside functions. As Reynolds points
 164

out [34], type abstraction and procedural abstraction are two orthogonal ways
to protect data.
Another advantage of the approach described in this section over the in-
strumentation of writes described in section 5.1 is that it often leads to fewer
run-time checks. In particular, checks at coercions can sometimes be proven re-
dundant and therefore can be eliminated. Consider the following function that
adds a . old suffix to a file name:

Af . OiffiienameCf i l e n a i n e ( c o n c a t ( s t r i i i g ( f ) , " . o l d " ) ) )

With the definition of P y (filename) given in section 4.2, it is easy to show that
if / belongs to PV(filename), then so does the concatenation of / and .old.
Hence, the OK test can be removed.
Of course, not all run-time tests can be removed this way: consider what
happens if the suffix is given as argument:

Af . As. OiiTfiienaneCf i l e n a m e ( c o n c a t ( s t r i n g ( f ) , s ) ) )

and the applet passes a suffix s starting with ". . /".

5.3 Instrumenting coercions without type abstraction

In some cases, the types t 6 Dom(Py) cannot be made abstract in the applet,
e.g. because it would make writing the applet too inconvenient, or entail too
much run-time overhead. We can adapt the approach presented in section 5.2
to these cases, by reverting to procedural abstraction and putting checks not
only at coercions, but also on all values of types t E Dom{PV) that come from
the applet. (This matches current practice in Unix kernels, where parameters
to system calls are always checked for validity on entrance to the system call.)
Figure 7 depicts this approach.
The checking of values coming from the applet is achieved by a standard
wrapping scheme applied to all functions of the execution environment, inserting
OKt coercions at all negative occurrences of types t £ Dom(Py). For instance,
if the execution environment needs to export a function / ; i —> t, it will actually
export the function Xx.f{OKt{x)), which validates its argument before passing
it to the original function.
We formalize these ideas in a slightly different way, in order to build upon
the results of section 5.2. Start from an applet environment defined by top-level
bindings of the form
l e t fi -.Ti = ai
We assume given a set TD' of type definitions against which the a, and the
applets are type-checked, and a valuation PV assigning permitted values to
named types in TD'.
We first associate a new named type i to each sensitive type t € D o m ( P y ) .
The type i is defined as synonymous with t, and is intended to represent those
 165

values of type t that have passed run-time validation. We define the i types by
taking
TD=TD'®[iy-^t\tG Dom{PV')]
and restrict the values they can take using the vaJuation PV defined by PV{i) =
PV'{t) and PV undefined on other types.
Let S be the substitution {i •*- t | t e Dom(PV")}. We transform the
bindings for the applet environment as follows:

l e t fi:n = W+{IC{E{ai)) : r^)

That is, we rewrite the terms aj to use the type i instead of t for all t €
Doxn{PV'); then apply the IC instrumentation scheme to it, thus adding an
OKf check to each coercion i{a); finally, apply the W"^ wrapping scheme to the
instrumented term, in order to perform both validation and coercion from t to i
on entrance, and the reverse coercion from f to i on exit. Wrapping is directed
by the expected type for its result, and is contravariant with respect to function
types. We thus define both a wrapping scheme W~^ for positive occurrences of
types and another W~ for negative occurrences.

W+{a:i) =zW-{a:L) = a
W+{a : t) = t{a) if t € Dom(PV^)
W-{a : t) = OKi{i{a)) if t E Dom(Py)
W^{a •f)=W-{a:t) = a if t ^ Dom(Py)
W+{a : Ti X ra) = (H^+(fst(a) : n ) , W^+(snd(a) : T^)
W-{a : n X ra) = (VK-(fst(a) : n ) , W-(snd(a) : T2)
W+{a : Ti ^ T2) == Xx.W+{a{W-{x : n ) ) : T2)
W~{a : n -^ T-i) = Xx.W-ia{W+ix : n ) ) : T2)
W+{a : T ref) = W~{a : r ref) = a
if no i € Dom{PV) occurs in T

Wrapping is not defined on reference types containing a. t G 'Dom{PV) because

there is no way to validate these references so that they are protected against
future modifications.
It is easy to see that the transformed bindings are well-typed: I!{ai) has type
S{Ti); the IC instrumentation preserves typing; the W'^ wrapping applied to a
term of type S{Ti) returns a term of type TJ, as shown by a simple induction
over Ti.
Moreover, the right-hand sides of the bindings always perform an OK^ check
before each coercion to a type i: this is ensured by the IC instrumentation for
the coercions initially in S{ai), and by definition of the wrapping scheme for the
coercions introduced by the wrapping.
Finally, the applets themselves are still type-checked in the original set TD'
of type definitions, in which the types i are not defined, and thus abstract for
the applet.
 166

We are therefore back to the situation studied in section 5.2: the types i
are abstract in the applets and all coercions to i are instrumented in the applet
environment. Thus, by Property 3, we obtain that the values of references with
types i ref always remain within PV{t.) = PV'(t) during the execution of any
well-typed applet.

Security property 4 Let e be the execution environment for applets and s the
initial store. Assume that e and s are obtained by evaluating a set of transformed
bindings l e t fi : n = W'^{IC{S{ai)) : Ti) as described above. Assume PV \=
s : S and PV, S \= e : E. Then, for every applet a well-typed in E and in the
initial set TD' of type definitions, we have Prot{PV,S),s,e \~ a -/> e r r .

6 Connections with object-oriented languages

Although the language used for this work is functional, the techniques developed
here translate reasonably well to object-oriented languages. Procedural abstrac-
tion as presented in section 3.2 corresponds most closely to Smalltalk-style pri-
vate instance variables: just as variables in a closure environment can only be
accessed by the code associated with that closure, private instance variables of a
Smalltalk object can only be accessed by the methods of that object. Java does
not offer a strictly equivalent mechanism: the p r i v a t e modifier makes instance
variables accessible not only to the methods of the object, but also to methods
of other objects of the same class. Still, the visibility rules associated with the
package mechanism and the p r i v a t e and p r o t e c t e d modifiers also ensure some
degree of procedural abstraction, since they restrict the set of methods that can
access a given instance variable.
Similarly, type abstraction as exploited in section 5.3 corresponds most
closely to f i n a l classes in Java. A f i n a l class containing only p r i v a t e fields
and no default constructors offers the same level of guarantees as our abstract
types: the applet cannot tamper with the fields of an existing object, nor create
an object with arbitrary initial values for the fields. In non-final classes, the
visibility rules might still ensure some degree of type abstraction, though it is
unclear how much is guaranteed. Subtype polymorphism in systems such as F^-,
[11] also provide some amount of type abstraction, but this is not so in Java
because the actual type of an object can be tested at run-time (downcasts).

7 Application: safety in t h e MMM browser

MMM [35] is a Web browser with applets developed by the second author. MMM
ensures safe execution of applets using various techniques similar in spirit to
those formulated earlier in this paper. (Essentially identical techniques axe now
used in the Switch Ware project at U. Penn [4,5], an active network infrastruc-
ture that allows safe downloading of applets on network routers.) Indeed, the
formalization presented in this paper grew out of a desire to make more sys-
tematic and prove correct the security techniques used in MMM. This section
 167

presents the main techniques used in the MMM safe execution environment for
applets, and relates them with the formal results we have obtained in this paper.

7.1 Type-safe dynamic linking

The MMM browser is written in Objective Caml [24] and compiled to bytecode
by the Caml bytecode compiler. The bytecode is then executed by the Caml
virtual machine. Applets are also compiled to Caml bytecode, then loaded in
memory and linked with the browser by the Caml dynamic linker (the Dynlink
library).
The dynamic linker can be configured to restrict the set of external modules
and C primitives that the object file can reference. MMM uses this feature to
enforce lexical scoping for applets: only the modules of MMM that comprise
the applet execution environment can be referenced by the applet object code;
modules internal to MMM are not made accessible to the applet.
The Caml bytecode was designed before applets were fashionable, and is
therefore ill-suited to JVM-style bytecode verification. In particular, a number
of low-level optimizations are performed in this bytecode (such as erasing type
annotations and conflating several high-level types into the same machine-level
representations) that makes type reconstruction on bytecode if not impossible,
but at least very difficult.
Instead of bytecode verification, the Objective Caml dynamic linker relies
on type annotations on the relocation information contained in object files to
guarantee type correctness. Bytecode object files contain a list of names of ex-
ternal modules referenced in the file, along with the type interfaces of those
modules. The object file thus records which interfaces for those external mod-
ules were used for type-checking the source code of the file. The dynamic linker,
then, checks that those interfaces are identical to those of the implementations
of those modules provided by the MMM browser. This ensures that the applet
and the browser have been type-checked against the same type specifications for
the applet-browser interface, and can therefore safely be linked together.
To keep them small, object files actually do not contain the whole Caml type
interfaces for their imported modules, but only MD5 checksums of those inter-
faces. The dynamic linker then compares interfaces by comparing their check-
sums. Since MD5 is a hash function of cryptographic quality, the probability
that two different interfaces have the same checksum is extremely low, and thus
we can safely assume that if the checksums are equal, then so are the interfaces.
However, this approach forces the two interfaces (the one expected by the object
file and the one provided by the MMMimplementation) to be exactly identical,
while, if we kept whole interfaces, we could use a more lenient comparison such
as Caml's interface matching. The latter would allow more flexibility in evolving
the applet-browser API.
Of course, comparing import interfaces at link-time guarantees type safety
only if the object code contained in the object file is indeed type-correct with
respect to the interfaces contained in the object file. Thus, we need to make sure
that this object file has been produced by a correct Caml compiler (a compiler
 168

that performs sound static type-checking and records correctly all external mod-
ules referenced) and has not been modified afterwards (e.g. by hand-modifying
some of the interfaces in the object file). This can be enforced in two ways in
the context of Web applets.
The first way is to transmit applets over the network as Caml source code,
which is type-checked and compiled locally (by calling the Objective Caml byte-
code compiler), then dynamically linked inside the browser. Unless the local host
is compromised, no tampering with the object files nor the Caml type-checker
itself can happen. Unlike applet systems based on source-level interpretation, we
still benefit from the efficiency of the Caml bytecode interpreter.
Transmission of applets in source form is often criticized on several grounds:
the source code is larger than compiled bytecode; local compilation takes too
much time; applet writers do not want to publicize their source. Our experience
with Caml is that bytecode object files are about the same size as the source
code (unless heavily commented); the bytecode compiler is very fast and the
compilation times are small compared with Internet latencies; finally, bytecode
is easy to decompile and does not offer significant protection against reverse
engineering.
Another way to ensure the correctness of type annotations in bytecode object
files is to rely on a cryptographic signature on the object file, which is checked
locally by the M M M browser against a list of trusted signers before linking the
object file in memory. Unlike Microsoft-style applet signing, this signature is
not necessarily made by the author of the applet, and carries no guarantees on
what the applet actually does; instead, the signature is made by the person or
site who performs the compilation and type-checking of the applet, and certifies
only that the applet passed type-checking by an unmodified Caml compiler and
that its object file has not been tampered since.
We initially envisioned having some centralized type-checking authority: one
or several reputable sites (such as INRIA) that accept source code from applet
developers, type-check and compile them locally, sign the object file and return it
to the applet developer. The object file can then be made available on the Web.
Seeing the authority's signature, browsers can trust the type information con-
tained in the object file and use it to validate the applet against the environment
they provide. In retrospect, this approach relies too much on a centralized au-
thority to scale up to the world-wide Web. However, it is perfectly suited to the
distribution of compiled applets across a restricted network such as a corporate
Intranet.

7.2 The execution environment for applets

The execution environment made available to applets is composed of a number
of modules that provide a safe subset of the Caml standard library, as well
as access to the CamlTk graphical user interface and to selected parts of the
MMM browser internals. In particular, an MMM applet can not only interact
graphically with the user and trigger navigation functions, like Java applets,
but also extend the MMM browser itself by installing new navigation functions,
 169

menu items, viewers for new types of embedded documents, display functions
for new HTML tags, and decoding functions for new "content-encoding" types.
The applet environment is derived from the OCaml standard library mod-
ules and the MMM implementation modules by two major techniques: hiding
of unsafe functions via module thinning, and wrapping of unsafe functions with
capability checks.

Module thinning: The ML module system offers the ability to take a restricted
view of an existing module via a signature constraint:
module RestrM = (M : RestrSig)
Only those components of M that are mentioned in the signature RestrSig are
visible in RestrM, and they have the types specified in RestrSig, which may
be less precise (more abstract) than their original types in M. This module thin-
ning mechanism thus supports both hiding components (functions, variables,
types, exceptions, sub-modules) of a module and making some type components
abstract. No code duplication occurs during thinning: the functions in RestrM
share their code with the corresponding functions in M.
Large parts of the applet environment are obtained by thinning existing
OCaml library modules. In the OCaml standard library, we hide by thinning all
file input-output functions, as well as related system interface functions (such
as reading environment variables and executing shell commands), and of course
all type-unsafe operations (such as array accesses without bound checks and
functions that operate on the low-level representation of data structures). For
good measure, we also make abstract a few data types such as lexer buffers to
hide their internal structure.
In the CamlTk GUI toolkit, we hide all functions that return widgets that
may belong to the browser or to other applets. Such functions include finding
the parent of a widget, finding a widget by its name, finding which widget owns
the focus or the selection, etc. This is less restrictive than checking that widgets
manipulated by the applet are children of its top-level widget: the applet can still
open new windows and populate them with widgets unrelated with its initial top-
level widget. Other functions that might affect browser widgets (such as binding
events on all widgets of a given class or tag) are also removed.

Capabilities: MMM uses capabilities to control potentially harmful function-

alities, such as file and network input/output, as well as extending the browser.
Individual capabilities include:
- Reading files whose names match a given regular expression.
— Writing files whose names match a given regular expression.
- Accessing the contents of Web documents (e.g. the pages being displayed).
— Installing extension functions in the browser.
Applets are given an initial set of capabilities at load time, which is empty for
applets loaded through the network, but allows document access and browser
 170

extension for applets loaded from the local disk. Capabilities are of course rep-
resented by an abstract data type, to prevent an applet from forging extra ca-
pabilities.
All input/output operations as well as registration of browser extensions
check the capabilities presented by the applet. If the applet does not possess the
capability to perform the requested operation, the browser prompts the user via
a pop-up window. The user can then refuse the operation (aborting the execution
of the applet), grant permission for this particular operation, or grant permission
for further operations of the same kind as well. In the latter case, the browser
extends in place the capabilities of the applet. To minimize the number of times
the user is prompted, an applet can also request in advance the capabilities it
needs later.

Hiding capability arguments: A well-known problem with capability-based

security for applets is that all sensitive functions require an extra capability
argument [42]. This clutters the source code for the applet and makes it difficult
to convert a piece of stand-alone code into an applet: capability arguments must
be added at many different places in the source code.
In functional languages, partial application can be used to pass the ca-
pability argument only once per sensitive function. Assume defined the func-
tions open_in_cap : c a p a b i l i t y —> s t r i n g —> in_channel and open_out_cap :
c a p a b i l i t y -> s t r i n g -4 out.channel as capability-checking variants of the
normal file opening functions open.in : s t r i n g —^ in.channel and open-Out :
s t r i n g -^ out.channel. Then, an applet can begin with the following defini-
tions:

let mycapa = C a p a b i l i t i e s . g e t ( )
let open_in = open_in_capa mycapa
let open_out = open_out_capa mycapa
... code using open_in and open_out as usual...

The remainder of the applet can then use the functions open_in and open_out
thus obtained by partial application as if they were the normal file opening
functions.
MMM makes this approach more convenient by using ML functors to per-
form the partiaJ applications. Functors are parameterized modules, presented
as functions from modules to modules. They provide a convenient mechanism
for parameterization en masse: rather than partially applying n functions to m
parameters, a structure containing the m parameters is passed once to a functor
that return a structure containing the n functions already partially applied to the
parameters. In the case of MMM, we thus have a number of functors that take
the applet's capability as argument and return structures defining capability-
enabled variants of the standard I/O and browser interface functions, with the
same interface as for standalone programs. For instance, to perform file I/O, an
MMM applet does the following:
 171

module MyCapa =
struct
let capabilities = Capabilities.get()
end
module 10 = Safeio(MyCapa)
open 10
. . . code using open_in and other I/O operations as usual...

7.3 Assessment

The MMM security architecture relies heavily on the three basic ingredients that
we considered in our formalization:

- Lexical scoping ensures that applets cannot access ail the modules composing
the browser and the Caml standard library, but only those "safe" modules
made available to the applet during linking.
- Type abstraction prevents the applet from forging or tampering with its
capability list. It also ensures that the applet cannot forge file descriptors or
GUI widgets, but has to go through the safe libraries to create them.
- Procedural abstraction is used systematically to wrap sensitive functions
(such as opening files or network connections, as well a.s installing a browser
extension) with capability checks.

In particular, capability checks inside environment functions are placed essen-

tially as suggested by the wrapping scheme W+ of section 5.3. In the MMM
implementation, those capability checks were placed by hand and the same type
was used for checked and unchecked data, instead of exploiting two different
types as in section 5.3. One of the motivations for our formalization was to sys-
tematize the placement of those capability checks, based on typing information.
Some aspects of the MMM security architecture are not accounted for by
our formalism. The first one is the necessity of taking copies of mutable objects
before validating them. A prime example is character strings, which, in Caml,
can be modified in place like character arrays. An applet could pass a string
containing a valid file name to the file opening function, then modify the string
in place between the time it is validated and the time the file is actually opened.
This concurrent modification can be achieved via multiple threads or via GUI
callback functions.
To avoid this attack, the execution environment must first take private copies
of all strings provided by the applet, then validate and utilize the private copies of
those strings. Symmetrically, the execution environment must not return strings
shared with its internal data structures to the applet, but copies of those strings.
Inserting those string copy operations is tedious and error-prone. ^Prom the
standpoint of applet security, it would be much better to have immutable strings
in the language. The guarantees offered by immutable strings could possibly be
achieved just by removing all string modification primitives from the applet
execution environment, but we have not investigated this approach yet.
 172

Another issue not addressed by our framework is validation of functional val-

ues, such as document decoders and HTML tag handling functions installed by
the applet. The main reason our framework does not handle function validation
is that membership of a functional value in a set PV{t) is in general undecid-
able. In the particular case of MMM, this problem does not arise: depending on
the capabilities given to the applet, either all functions presented by the applet
are rejected, or they are all accepted. For more complex situations, we could
also move the validation inside the applet-provided function, by wrapping this
function with checks on its inputs or outputs.

Resources that can be explicitly deallocated and reassigned later raise in-
teresting problems. A prime example is Unix file descriptors, which are small
integers. The type of file descriptors is abstract, and an applet cannot forge a
file descriptor itself. However, it could open a file descriptor on a permitted file,
close it immediately, keep the abstract value representing the descriptor, and
wait until the browser opens a file or network connection and receives the same
file descriptor in return. Then, the applet can do input/output directly on the
file descriptor, thus accessing unauthorized files or network connections. This
"reuse" attack is of course possible only with explicitly-deallocated resources:
with implicit deallocation, the resources cannot be deallocated and reallocated
as long as the applet keeps a handle on the resource. MMM addresses this prob-
lem by wrapping Unix file descriptors in an opaque data structure containing
a "valid" bit that is set to false when the file descriptor is closed, and checked
before every input/output operations.

The last MMM security feature not addressed by our framework is confiden-
tiality. Our framework focuses on ensuring the integrity of the browser and of
the host machine. The MMM applet environment contains security restrictions
to ensure integrity (such as controlled write access to files), but also restrictions
intended to protect the confidentiality of user data (such as controlled read ac-
cesses to files and restricted access to the network). Some restrictions address
both integrity and confidentiality problems: for instance, a malicious document
decoder could both distort the document as displayed by the browser (an in-
tegrity threat) and leak confidential information contained in the document to
a third party (a confidentiality threat).

However, the confidentiality policy enforced by the MMM applet environ-

ment is simple and does not go beyond traditional access control. While the
formal framework presented in this paper focuses on integrity via access control
on writes, we believe that parts of this framework (procedural encapsulation
and type abstraction) can be extended to deal with access control on reads,
thus ensuring simple confidentiality properties adequate for MMM. More ad-
vanced confidentiality policies (such as allowing an applet to read local files or
open network connections, but not both) are definitely beyond the scope of our
framework, however.
 173

8 Related work
8.1 Type systems for security
The work most closely related to ours is the recent formulations of Denning's
information flow approach to security [13,14] as non-standard type systems by
Palsberg and 0rbaek [31], Volpano and Smith [41,40], and Heintze and Riecke
[20]. (Abadi et al. [2] reformulate some of those type systems in terms of a more
basic calculus of dependency.) The main points of comparison with our work are
listed below.
Information flow vs. integrity: The type systems developed in previous works
all focus on secrecy properties, following the information flow approach. In par-
ticular, they allow high-security data to be exposed as long as no low-security
code uses this data. Our work focuses on more basic integrity properties via ac-
cess control. We view those integrity guarantees as a prerequisite to establishing
meaningful confidentiality properties.
Imperative vs. purely functional programs: [31] and [20] consider purely
functional languages in the style of the A-calculus. This makes formulating the
security properties delicate: [31] proves no security property properly speak-
ing, only a subject reduction property that shows the internal consistency of the
calculus, but not its relevance to security; [20] does show a non-interference prop-
erty (that the value of a low-security expression is independent of the values of
high-security parameters), but it is not obvious how this result applies to actual
applet/browser interactions, especially input/output. Instead, we have followed
[41,40] and formulated our security policy in terms of in-place modifications on
a store, which provides a simple and intuitive notion of security violation.
Run-time validation of data: Only [31] and our work consider the possibility
of checking low-security data at run-time and promoting them to high security. In
[41,40,20], once some data is labeled "low security", it remains so throughout
the program and causes all data it comes in contact with to be marked "low
security" as well. We believe that, in a typical applet/browser interaction, this
policy leads to rejecting almost all applets as insecure. Run-time validation of
untrusted data is essential in practice to allow a reasonable range of applets to
run.
Subtyping vs. named types and coercions: All previous works consider
type systems with subtyping, which provides a good match for the flow analysis
approach they follow [30]. In contrast, we only use type synonyms with possibly
checked coercions between a named type and its implementation type. However,
the connections between subtyping and explicit coercions are well known [9], and
we do not think this makes a major difference.

8.2 Security of applets

Concerning the security issues raised by applets in general, we are aware of case
studies of security flaws [12], as well as informal descriptions of current and
 174

proposed security architectures for applets [16,44,35,42,18]. Our work seems to

be one of the first formal studies of applet security.
On the Java side, considerable effort has been expended in proving the sound-
ness of the Java type system [15,37,29] and of the JVM bytecode verifier [33,
36,17]. Other aspects of Java that are equally important for security, such as
formalizing the visibility rules and the encapsulation guarantees they provide,
have only recently started to receive attention [22]. We are not aware of any
formal description of security policies for Java applets.
Proof-carrying code [28,27] provides an elegant framework to establish the
safety of mobile code, but requires general proof from the applet's developer.
Our approach lies at the other end of the complexity spectrum: all we require
from the applet is that it is well typed in a simple, standard type system.

9 Concluding remarks

We have identified three basic techniques for enforcing a fairly realistic security
policy for applets: lexical scoping, procedural abstraction, and type abstraction.
These programming techniques are of course well known, but we believe that this
work is the first to characterize precisely their implications for program security.
The techniques proposed here seem to match relatively well current practice
in the area of Web applets. In particular, they account fairly well for Rouaix's
implementation of safe libraries in the MMM browser.
Our techniques put almost no constraints on the applets, except being well-
typed in a simple, completely standard type system. The security effort is con-
centrated on the execution environment provided by the browser. Typing the
applets in a richer type system, such as the type systems for information flow of
[31,41,40,20] or the effect and region system of [38], could provide more infor-
mation on the behavior of the applet and enable more flexible security policies
in the execution environment. However, it is impractical to rely on rich type
systems for applets, because these type systems are not likely to be widely ac-
cepted by applet developers. Whether these rich type systems can be applied to
the execution environment only, while still using a standard type system for the
applets, is an interesting open question.
On the technical side, the proofs of the type-based security properties are
variants of usual type soundness proofs. It would be interesting to investigate
the security content of other classical semantic results such as representation
independence and logical relations. Given the importance of communications
between the applet and its environment, it could be worthwhile to reformulate
our security results for a calculus of communicating processes [6,3,1].

Acknowledgements

This work has been partially supported by GIE Dyade under the "Verified In-
ternet Protocols" project. A preliminary version of this paper appeared in the
 175

proceedings of the 25th ACM symposium on Principles of P r o g r a m m i n g Lan-

guages.

References
1. M. Abadi. Secrecy by typing in security protocols. In Theoretical Aspects of
Computer Software '97, volume 1281 of Lecture Notes in Computer Science, pages
611-638. Springer-Verlag, Sept. 1997.
2. M. Abadi, A. Banerjee, N. Heintze, and J. G. Riecke. A core calculus of dependency.
In 26th symposium Principles of Programming Languages, pages 147-160. ACM
Press, 1999.
3. M. Abadi and A. D. Gordon. Reasoning about cryptographic protocols in the Spi
calculus. In CONCUR'97: Concurrency Theory, volume 1243 oi Lecture Notes in
Computer Science, pages 59-73. Springer-Verlag, July 1997.
4. D. S. Alexander, W. A. Arbaugh, M. W. Hicks, P. Kakkar, A. D. Keromytis, J. T.
Moore, C. A. Gunter, S. M. Nettles, and J. M. Smith. The SwitchWare active
network architecture. IEEE Network, 12(3):29-36, 1998.
5. D. S. Alexander, W. A. Arbaugh, A. D. Keromytis, and J. M. Smith. Security in
active networks. In J. Vitek and C. Jensen, editors, Secure Internet Programming,
Lecture Notes in Computer Science. Springer-Verlag Inc., New York, NY, USA,
1999.
6. J.-P. Banatre and C. Bryce. A security proof system for networks of communicating
processes. Research report 2042, INRIA, Sept. 1993.
7. J.-P. Billon. Security breaches in the JDK 1.1 beta2 security API. Dyade, h t t p : /
/www.dyade.fr/fr/actions/VIP/SecHole.html, Jan. 1997.
8. N. S. Borenstein. Email with a mind of its own: the Safe-Tel language for en-
abled mail. In IFIP International Working Conference on Upper Layer Protocols,
Architectures and Applications, 1994.
9. V. Breazu-Tannen, T. Coquand, C. A. Gunter, and A. Scedrov. Inheritance as
implicit coercion. Information and Computation, 93(1):172-221, 1991.
10. K. Brunnstein. Hostile ActiveX control demonstrated. RISKS Forum, 18(82), Feb.
1997.
11. L. Cardelli, S. Martini, J. C. Mitchell, and A. Scedrov. An extension of system F
with subtyping. Information and Computation, 109(l-2):4-56, 1994.
12. D. Dean, E. W. Felten, D. S. Wallach, and D. Balfanz. Java security: Web browsers
and beyond. In D. E. Denning and P. J. Denning, editors, Internet Besieged:
Countering Cyberspace Scofflaws, pages 241-269. ACM Press, 1997.
13. D. E. Denning. A lattice model of secure information flow. Commun. ACM,
19(5):236-242, 1976.
14. D. E. Denning and P. J. Denning. Certification of programs for secure information
flow. Commun. ACM, 20(7):504-513, 1977.
15. S. Drossopoulou and S. Eisenbach. Java ib type safe - probably. In Proc. 11th
European Conference on Object Oriented Programming, volume 1241 of Lecture
Notes in Computer Science, pages 389-418. Springer-Verlag, June 1997.
16. M. Erdos, B. Hartman, and M. Mueller. Security reference model for the Java
Developer's Kit 1.0.2. JavaSoft, http://java.sun.com/security/SRM.html, Nov.
1996.
17. S. N. Freund and J. C. Mitchell. A type system for object initialization in the
Java bytecode language. In Object-Oriented Programming Systems, Languages
and Applications 1998, pages 310-327. ACM Press, 1998.
 176

18. L. Gong. Java security axchitecture (JDKl.2). JavaSoft, h t t p : / / j a v a . s u n . c o m /

products/jdk/1.2/docs/guide/security/spec/security-spec.doc.html, Oct.
1998.
19. J. Gosling and H. McGilton. The Java language environment - a white paper.
JavaSoft, h t t p : / / j a v a . s u n . c o m / d o c s / w h i t e / l a n g e n v , May 1996.
20. N. Heintze and J. G. Riecke. The SLam calculus: programming with secrecy and
integrity. In 25th symposium Principles of Programm,ing Languages, pages 365-377.
ACM Press, 1998.
21. D. Hopwood. Java security bug (applets can load native methods). RISKS Forum,,
17(83), Mar. 1996.
22. T. Jensen, D. Le Metayer, and T. Thorn. Security ajid dynamic class loading in
Java: A formalisation. In International Conference on Computer Languages 1998,
pages 4-15. IEEE Computer Society Press, 1998.
23. X. Leroy. Polymorphic typing of an algorithmic language. Research report 1778,
INRIA, 1992.
24. X. Leroy, J. Vouillon, D. Doligez, et al. The Objective Caml system. Software and
documentation available on the Web, h t t p : / / c a m l . i n r i a . f r / o c a m l / , 1996.
25. R. Milner, M. Tofte, R. Harper, and D. MacQueen. The definition of Standard ML
(revised). The MIT Press, 1997.
26. G. Morrisett, M. Pelleisen, and R. Harper. Abstract models of memory manage-
ment. In Functional Programming Languages and Computer Architecture 1995,
pages 66-77. ACM Press, 1995.
27. G. C. Necula. Proof-carrying code. In 24th symposium Principles of Programming
Languages, pages 106-119. ACM Press, 1997.
28. G. C. Necula and P. Lee. Safe kernel extensions without run-time checking. In
Proc. Symp. Operating Systems Design and Implementation, pages 229-243. Usenix
association, 1996.
29. T. Nipkow and D. von Oheimb. JavaLight is type-safe — definitely. In 25th sym-
posium Principles of Programming Languages, pages 161-170. ACM Press, 1998.
30. J. Palsberg and P. O'Keefe. A type system equivalent to flow analysis. ACM TVans.
Prog. Lang. Syst, 17(4):576-599, 1995.
31. J. Palsberg and P. 0rbaek. Trust in the A-calculus. Journal of Functional Pro-
gramming, 7(6):557-591, 1997.
32. G. D. Plotkin. A structural approach to operational semantics. Technical Report
DAIMI FN-19, Aarhus University, 1981.
33. Z. Qian. A formal specification of a large subset of Java Virtual Machine instruc-
tions. In J. Alves-Foss, editor. Formal Syntax and Semantics of Java, Lecture
Notes in Computer Science. Springer-Verlag, 1998. To appear.
34. J. C. Reynolds. User-defined types and procedural data structures as comple-
mentary approaches to data abstraction. In C. Gunter and J. Mitchell, editors.
Theoretical aspects of object-oriented programming, pages 13-23. MIT Press, 1994.
35. F. Rouaix. A Web navigator with applets in Caml. In Proceedings of the 5th
International World Wide Web Conference, Computer Networks and Telecommu-
nications Networking, volume 28, pages 1365-1371. Elsevier, May 1996.
36. R. Stata and M. Abadi. A type system for Java bytecode subroutines. In 25th sym-
posium Principles of Programming Languages, pages 149-160. ACM Press, 1998.
37. D. Syme. Proving JavaS type soundness. Technical Report 427, University of
Cambridge Computer Laboratory, June 1997.
38. J.-P. Talpin and P. Jouvelot. The type and effect discipline. Information and
Computation, lll(2):245-296, 1994.
 177

39. M. Tofte. Type inference for polymorphic references. Information and Computa-
tion, 89(1), 1990.
40. D. Volpano and G. Smith. A type-based approach to program security. In Proceed-
ings of TAPSOFT'97, Colloquium on Formal Approaches in Software Engineer-
ing, volume 1214 of Lecture Notes m Computer Science, pages 607-621. Springer-
Verlag, 1997.
41. D. Volpano, G. Smith, and C. Irvine. A sound type system for secure flow analysis.
Journal of Computer Security, 4(3):1-21, 1996.
42. D. S. Wcillach, D. Balfanz, D. Dean, and E. W. Felten. Extensible security ar-
chitectures for Java. Technical report 546-97, Department of Computer Science,
Princeton University, Apr. 1997.
43. D. S. Wallach and E. W. Felten. Understanding Java stack inspection. In Pro-
ceedings of the 1998 IEEE Symposium on Security and Privacy. IEEE Computer
Society Press, 1998.
44. F. Yellin. Low level security in Java. In Proceedings of the Fourth International
World Wide Web Conference, pages 369-379. O'Reilly, 1995.

A Appendix: the containment lemma

In this appendix, we formalize the intuition t h a t if the type of the applet environ-
ment does not contain certain r e f types, then references of those types cannot
be exchanged between the applet and the environment, and remain "contained"
in one of them.
We a n n o t a t e each source-language term a as coming either from t h e execution
environment {uenv) or from the applet {aapp)- We let m,n range over the two
"worlds" env and app, and write m for the complement of m, i.e. emJ = app and
'app = env.
Let T be a set of type expressions satisfying t h e following closure property:
if r € T , then all types r ' t h a t contain T as a sub-term also belong to T. (In
section 5.1, we take T to be the set of all types containing an occurrence of
t ref; in appendix B, T is the set of all types containing an occurrence of any
r e f type). We partition the set of locations into three countable sets:

- Lapp is the set of locations with type r r e f G T t h a t have been allocated by

the applet;
- Lenv is t h e set of locations with type r r e f e T t h a t have been allocated by
the environment (i.e. either initially present in the applet environment, or
allocated by environment functions);
- Lshared IS t h e Set of locations whose type r r e f does not belong to T.

To ensure t h a t locations allocated during evaluation are drawn from the correct
set, we assume all source terms aj„ annotated with their static type r and their
world m and replace the evaluation rule for reference creation (rule 8) by the
following rule:
 178

>p,e,s\- al^ -> vjs' t ^ Dom(s') U T)om.{ip)

i G L^nv if T ref G T and m = env
(• € -^app if T ref e T and m = app /gi-j
(• e Lshared if T Tef ^ T

ip,e,s\- ref ( a ) ; , " ' -f £/s'{^ ^ ?;}

It is easy to see that the modified rule produces the same evaluation derivations
as the initial rule, up to a renaming of fresh locations: ii ip,e,s \- a ^ r with the
initial rule, then ifi,e,s \- a -^ r' with the modified rule, where r' is identical to
r up to a renaming of locations not in Dom(s) U Dom{(p).
We say that a value u in a store s is contained in world m, and we write
Cmiv, s), if any source term operating on u in s can directly access only locations
that are in Lm or Lshared, but not locations in L^- Formally, Cm{v, s) is defined
by case analysis on v, as follows:
- Cm{b,s) is always true
- C„j(Aa;.a„[e],s) if C„(e,s)
- Cm{{vi,V2),s) if Cm(wi,s) and Cmiv2,s)
- C^{e, s)iil^ L™ and C,„(s(£), s)
- Cm(e,s) if Cm{e{x),s) for all x 6 Dom(e).
Notice that for closures, it's the world n of the function body a„ that deter-
mines the containment of the closure environment, not the world m in which
the closure value is being used. The reason is that the environment values can
only be accessed by the function body, not arbitrary terms of world m. This is
characteristic of procedural abstraction in the sense of section 3.2.
As usual, the definition of C above is not well-founded by induction on v.
We view the equations above as fixpoint equations for an operator, which is
increasing, and define C as the greatest fixpoint of that operator. (The smallest
fixpoint is always false on value/store pairs that contain a cycle, which is not
what we want; it's the greatest fixpoint that gives the expected behavior for C.
See [39] for detailed explanations.)
Here are two important lemmas on the C predicate. First, assigning a value
contained in world m to a location which is not in Ljn preserves the containment
of all other values.
L e m m a 5. Assume either i G L^ and Cm(v,s), or i £ Lghared and Capp{v,s)
and Cenv{v,s). Then, for all values w and worlds n, C„(w,s) implies
Cniw,s{e^v}).
Proof. The proof is a standard argument by coinduction, close to the proof of
Lemma 4.6 in [39]. The only non-trival case is w = i (the modified location).
By hypothesis C„(w,s), we have i ^ Ljc and C„(s(^),s). Write s' = s{£ •«- v}.
Notice that s'{i) = v. If n = m, we have C„(u, s) by assumption, hence C„(w, s')
by the coinduction hypothesis, from which it follows that C„(£, s'). If n = m,
since i ^ Ln, we have i ^ L„i and thus it must be the case that £ £ Lshared and
Capp(v,s) and Ceni)(w,s) to comply with the assumptions of the lemma. Thus,
we have C„(v, s) and we conclude as in the previous case.
 179

Second, values that belong to a type T ^ T can be exchanged between the

app and env worlds without breaking containment.

Lemma 6. Assume S \= v : T and \^ s : S. If T ^ T, then Cmii>,s) implies

Cm{v, s) for all worlds m.

Proof. By structural induction on r. If r is a base type or a function type, then

f is a base value or a closure, and the containment of v is independent of the
world m. If r is a product type ri x TJ, the closure condition over T guarantees
that Ti ^ T and T-Z ^ T; the result follows from the induction hypothesis. Finally,
if T is a ref type, we have v = i and r = S{i) ref. Since r ^ T, it follows that
i belongs to Lghared, but neither to Lapp nor Lenv Hence, I ^ Lm- Moreover,
5(£) ^ T by the closure condition on T, hence Cm{s{£),s) by application of the
induction hypothesis. It follows that Cm{£,s).

We can now show that containment is preserved at each evaluation step:

Lemma 7 (Containment lemma). Let Eapi be the type of the execution envi-
ronment for applets. Assume Eapi{x) ^ T for all x € Dom(jEapi). Further assume
E h am '• T and S \= e : E and \= s : S and Cm{e,s). If (f,e,s h Om —> v/s',
then Cm{v,s'), and for all values iv and worlds n such that Cn{w,s), we have
Cn{w,s').

Proof. The proof is by induction on the evaluation derivation and case analysis
on a. Notice that by Lemma 2, we have the additional result that there exists a
store typing 5 ' extending S such that S' \= v : T and \= s' : S'. This makes the
semantic typing hypotheses go through the induction. The interesting cases are
assignment and function application; the other cases are straightforward.

Assignment: a is (af " ' := 0-2). We apply the induction hypothesis twice,
obtaining

ip,e,s\- ai -)• vi/si ip,e,si h 0-2 ->• V2/S2

Cmivi,S2) Cr,i{v2,S2)
Cn{w,s) implies C„(w,S2) for all n,w
52 1= Ui :CTref 52 [= ^'a : cr |= *2 : 52.

Hence, vi is a location £, and since v\ is contained in m, we have (. ^ L^.

Therefore, either^ e Lm or £ e Lgi^ared- But in the latter case, a ^ T by
construction of Lghared, hence Cm(f2: S2) implies Cm{v2,S2) as well by Lemma 6.
In both cases, the hypotheses of Lemma 5 are met. Writing s' = S2{1 <- V2},
we obtain the expected result: C„(w;,s) implies C„{w,s') for all n,w. The other
expected result, Cm((),*'')i is trivial.
 180

Application: a is a'[~^'^{a2)- By applying the induction hypothesis twice, we

obtain

Lp,e,s\- ai -^ vi/si ip,e,si\- a2 -^ V2/S2

Cm (W1,S2) Cm{v2,S2)
C„(w,s) implies C„{w,S2) for all n,iu
S2 \= Vi : a ^ T S2 \= V2 '• (T \= S2 '• 82-

Hence, vi is a closure Ax.a^[e'], and the last evaluation rule used is rule 4.
If n = m (intra-world call), we have Cm{e-'{x <- V2},S2) as a consequence
of Cm{vi,S2) and Cm.{v2,S2), and the two conclusions follows easily from the
induction hypothesis applied to the evaluation of a^.
If n = m (cross-world call), then the type <T —> r of the function must
occur as a sub-term of the typing environment Eapi- Hence, a ^ T and T ^ T.
Since the type of V2 is not in T, by Lemma 6 it follows that Cn{v2,S2)- Hence,
Cn{e'{x <— V2},S2), and we can apply the induction hypothesis to the evaluation
of a'„: tp, e'{x <— D2}, S2 ^' ojj -> v/s'. The resulting value v is contained in world
n and has type r; applying again Lemma 6, we get Cm{i>,s'), which is the
expected result.

B Appendix: modifiable locations in the case of

systematic procedural encapsulation
In this appendix, we formalize the notion of modifiable locations, as introduced
in section 3.2, in the particular case where the applet environment Capi is well-
typed and its type Eapi does not contain any ref types. This is not to say that
Bapi is purely functional: the functions it provide may very well have side-effects
and use references internally. Only, all those internal references must be encap-
sulated inside functions and never handed to the applet directly. This systematic
procedural encapsulation does not reduce expressiveness in any significant way,
and at any rate is a reasonable thing to do given that our goal is to characterize
semantically procedural encapsulation.
Prom the technical side, requiring that no r e l types occur in the applet's
interface Eapi has an interesting consequence: only source terms from the browser
can operate directly on locations allocated by the browser, and only terms from
the applet can operate directly on locations allocated by the applet. Thus, all
references are contained (in the sense of appendix A, taking T to be the set of all
types where the ref constructor occurs) either in the browser or in the applet,
but never go from one world to the other.
We then rely on the notion of containment to define the set ML{v, s) of loca-
tions modifiable from value v in store s as the smallest fixpoint of the following
equations:

ML{b, s) = 0
ML{{vi,V2),s) = ML{vi,s)UML{v2,s)
 181

ML{Xx.a[e],s) = {£ | there exists vi,si,V2,S2 such that

Cappivi,si) and si(£) = s{£) for all £ G ieni;, and either
{£ <- 0},e{a; -f- fi},Si h a -)^ e r r
or 0,e{a; 4- Wi},s h a -^ 1^2/^2
and £ 6 MZ,('y2, S2) U ML{eapi, S2)}
ML(e,s)= (J Mi(e(a;),s)
xeDom(e)

The case for closures follows conditions 1 and 4 in the informal discussion from
section 3.2. For condition 5, we use the condition Capp{vi,si) instead of the
more natural i ^ ML{vi, si), so that the equations remain increasing in ML and
the existence of the smallest fixpoint is guaranteed. The typing hypothesis (that
Eapi contains no ref types) renders condition 3 vacuous, and also dispenses us
with defining ML over locations.
The following lemma show that modifiable locations are indeed the only
locations modified during the appHcation of a closure.

Lemma 8. Letp be a set of locations, withp C Lenv Let Xx.aenv[e] be a closure

of an environment function. Assume p fl ML[Xx.a[e],s) = 0 and Capp{v,s). If
Prot{p),e{x <- v},s h a —> r, then r ^ err. Instead, r = v'/s', and moreover
p n ML(v', s') ~ 0 and p D ML{eapi, s') = 0.

Proof. Assume, by way of contradiction, that r = e r r . Given the evaluation

rules, there must exist i & p such that {£ <- 0},e{a; i- v},s \- a -^ e r r . By defi-
nition of ML, this means that £ G ML{Xx.a[e],s). This contradicts the hypothesis
pr\ML{Xx.a[e], s) — 0. Hence, r ^ err. We therefore have Prot{p),e{x <- v],s\-
a —^ v'/s' for some v' and s'. Given the evaluation rules, this implies that we
can also derive 0,e{a; <r- v},s \- a ^ v'/s'. Hence, by definition of ML, any £
belonging to ML(v',s') or ML{eapi,s') also belongs to ML{Xx.a[e],s). It follows
that p n ML{v',s') = ill a n d p n ML(eapi,s') = 0.

Using the notion of modifiable locations instead of reachable locations, we

finally obtain a security property similar to Property 1: the execution of an
applet cannot write to locations that are not modifiable from the initial execution
environment.

Security property 5 Assume S \= eapi • Eapi and [= s : 5 . Further assume

that Eapi contains no ref types. If p Q Dom(s) and pf) ML{eapi,s) = 0, then
for all applets a, we have Prot{p), Capi, s h a y^ err.

The property follows from the inductive lemma below.

Lemma 9. Assume S \= Capi : Eapi o-nd \= s : S and Eapi contains no ref

types. Further assume p C Le,^^ and pH ML[eapi,s) = 0. Assume E \- Uapp : r
and S \= e : E and Cappie, s). If Prot{p),e, s \- Uapp -> r, then r ^ err. Instead,
r = v'/s' and we have p D ML{eapi,s') = 0
 182

Proof. The proof is by induction on the evaluation derivation. We rely on Lem-

mas 2 and 7 to ensure that the semantic typing and containment hypotheses go
through the induction. The two non-obvious cases are assignment and applica-
tion of a closure of a browser function. We write (/? = Protijp).

Assignment: a is ax := a-z- Applying the induction hypothesis twice, we obtain

ip,e,s\- ai -f vi/si andv3,e,si h a-i -> ^^2/''''2, with Cnpp{vi,S2) and Capp{v2,S2)
and S2 \= vx : a ref and 52 |= wj - (^ and p Pi ML{eapi,S2) = 0. Hence, vi is a
location £, and since vi is contained in app, we have i £ Lapp- Thus, i ^ p and
the evaluation of the assignment does not result in e r r . Moreover, by definition
of ML, we have ML{eapi, s) = ML{eapi, s') if s{i) = s'{l) for all £ e Lenv In the
present case, the store s' at the end of the evaluation is S2{i •<— V2}, with E ^ Lenv
by containment, hence ML{eapi, s') = ML{eapi, S2) and thus pr\ML{eapi, s') = 0
as expected.

Application: a is 01(02). By induction hypothesis, we have ip,e,s\- ai -> Vy/si

and if,e,si 1- 02 -> V2/S2, with C„pp(ui,S2) and Cnpp(f2,S2) and 52 |= ui : cr -^
r and 52 \= V2 : cr and p D ML{eapi,S2) = 0- Hence, Vi is a closure \x.a'^[e'].
If m = app (the function comes from the applet), the result follows from the
induction hypothesis applied to the evaluation of a'. If m = env (the function
comes from the applet environment), the closure Xx.a'„^[e'] can only be obtained
by looking up a variable bound in Capi, then possibly performing some function
applications or f s t and snd operations. Thus, we have ML(Ax.a^^[e'],S2) C
ML{eapi, S2), and the result follows by Lemma 8.
 Part II

Concepts
 The Role of Trust Management in Distributed
Systems Security

M a t t Blaze, J o a n Feigenbaum, J o h n loannidis, and Angelos D. Keromytis

A b s t r a c t . Existing authorization mechanisms fail to provide powerful

and robust tools for handling security at the scale necessary for today's
Internet. These mechanisms are coming under increasing strain from the
development and deployment of systems that increase the programma-
bility of the Internet. Moreover, this "increased flexibility through pro-
grammability" trend seems to be accelerating with the advent of propos-
als such as Active Networking and Mobile Agents.
The trust-management approach to distributed-system security was de-
veloped as an answer to the inadequacy of traditional authorization
mechanisms. Trust-management engines avoid the need to resolve "iden-
tities" in an authorization decision. Instead, they express privileges and
restrictions in a programming language. This allows for increased flexi-
bility and expressibility, as well as standardization of modern, scalable
security mechanisms. Further advantages of the trust-management ap-
proach include proofs that requested transactions comply with local poli-
cies and system architectures that encourage developers and administra-
tors to consider an application's security policy carefully and specify it
explicitly.
In this paper, we examine existing authorization mechanisms and their
inadequacies. We introduce the concept of trust management, explain its
basic principles, and describe some existing trust-management engines,
including PoHcyMaker and KeyNote. We also report on our experience
using trust-management engines in several distributed-system applica-
tions.

1 Introduction
W i t h the advent of the Internet, distributed computing has become increasingly
prevalent. Recent developments in programming languages, coupled with t h e in-
crease in network bandwidth and end-node processing power, have m a d e t h e Web
a highly dynamic system. Virtually every user of the Internet is at least aware of
languages such as Java [15], JavaScript, Active-X, and so on. More "futuristic"
projects involve computers running almost exclusively downloaded interpreted-
language applications (Network P C ) , or on-the-fly programmable network infras-
tructures (Active Networks). On a more m u n d a n e level, an increasing number
of organizations use the Internet (or large Intranets) to connect their various
offices, branches, databases, etc.
All of these emerging systems have one thing in common: t h e need to grant or
restrict access to resources according to some security policy. There are several
issues worth noting.
 186

First, different systems and applications have different notions of what a re-
source is. For example, a web browser may consider CPU cycles, network band-
width, and perhaps private information to be resources. A database server's
notion of "resource" would include individual records. Similarly, a banking ap-
plication would equate resources with money and accounts. While most of these
resources can be viewed as combinations of more basic ones (such as CPU cy-
cles, I/O bandwidth, and memory), it is often more convenient to refer to those
combinations as single resources, abstracting from lower-level operations. Thus,
a generic security mechanism should be able to handle any number and type of
resources.
What should also be obvious from the few examples mentioned above is
that different applications have different access-granting or -restricting policies.
The criteria on which a decision is based may differ greatly among different
applications (or even between different instances of the same application). The
security mechanism should be able to handle those different criteria.
One security mechanism often used in operating systems is the Access Control
List (ACL). Briefly, an ACL is a list describing which access rights a principal has
on an object (resource). For example, an entry might read "User Foo can Read
File Bar." Such a list (or table) need not physically exist in one location but may
be distributed throughout the system. The t/nix^^-filesystem "permissions"
mechanism is essentially an ACL.
ACLs have been used in distributed systems, because they are conceptually
easy to grasp and because there is an extensive literature about them. How-
ever, there are a number of fundamental reasons that ACLs are inadequate for
distributed-system security, e.g.,

- Authentication: In an operating system, the identity of a principal is well

known. This is not so in a distributed system, where some form of authenti-
cation has to be performed before the decision to grant access can be made.
Typically, authentication is accomplished via a username/password mecha-
nism. Simple password-based protocols are inadequate in networked comput-
ing environments, however, even against unsophisticated adversaries; simple
eavesdropping can destroy security. Other recently developed mechanisms
include:
• One-Time passwords, which do not secure the rest of the session.
• Centralized ticket-based systems, such as Kerberos [22]. Problems with
such systems include the necessity for an authentication server (and for
frequent communication with it) and implicit trust assumptions.
• Public-key based authentication protocols, which are considered the "sta-
te of the art" for scalable authentication systems.
— Delegation is necessary for scalabihty of a distributed system. It enables de-
centralization of administrative tasks. Existing distributed-system security
mechanisms usually delegate directly to a "certified entity." In such systems,
policy (or authorizations) may only be specified at the last step in the dele-
gation chain (the entity enforcing policy), most commonly in the form of an
ACL. The implication is that high-level administrative authorities cannot
 187

directly specify overall security policy; rather, all they can do is "certify"
lower-level authorities. This authorization structure leads easily to inconsis-
tencies among locally-specified sub-policies.
- Expressibility and Extensibility: A generic security mechanism must be able
to handle new and diverse conditions and restrictions. The traditional ACL
approach has not provided sufficient expressibility or extensibility. Thus,
many security policy elements that are not directly expressible in ACL form
must be hard-coded into applications. This means that changes in security
policy often require reconfiguration, rebuilding, or even rewriting of applica-
tions.
- Local trust policy: The number of administrative entities in a distributed
system can be quite large. Each of these entities may have a different trust
model for different users and other entities. For example, system A may trust
system B to authenticate its users correctly, but not system C; on the other
hand, system B may trust system C. It follows that the security mechanism
should not enforce uniform and implicit policies and trust relations.

We believe that these points constitute a forceful argument that Authenti-

code, X.509, and, generally, the use of identity-based public-key systems in con-
junction with ACLs^ are inadequate solutions to distributed (and programmable)
system-security problems. Even modern ACL-based systems like DCE fall some-
what short of satisfyingly addressing the extensibility, expressibility, and dele-
gation issues, despite some work in these directions [10].
Furthermore, one finds insecure, inadequate, or non-scalable authentication
mechanisms (such as username/password, One-Time passwords, and hardware
token authentication) used in conjunction with ACLs. Finally, many pohcy prob-
lems are left unsolved by the "binary" authorization model employed widely in
Web security (and elsewhere): access is granted on the condition that the re-
questing principal has a certificate by a particular Certification Authority (CA).
We call this a binary authorization model because it means essentially "all-or-
nothing" access. While this is sufficient in a small number of cases (e.g., when
read-only access to a web page is the only decision that need be made), it is
obvious that this approach neither scales nor is extensible.
We believe that these unintuitive and in many ways problematic mechanisms
are in use simply because of the lack of alternatives that are better suited to
distributed systems. Developers have tried to adapt existing tools to their se-
curity model or vice versa, and neither strategy has worked particularly well.
Accordingly, misconfiguration and unexpected component interactions are caus-
ing security breaches and, inevitably, loss of confidence in the security tools.
We also beheve that the new Internet services now emerging as a result of
increased programmability will require powerful and expressive authorization

^ Sometimes the public keys are hardcoded into the apphcation, which deprives the
environment of even the hmited flexibility provided by ACLs. Such an example is
the under-development IEEE 1394 Digital Content Protection standard.
 188

mechanisms without the problems present in ACL-based systems^. Naturally,

existing services and protocols can also make good use of such mechanisms.
Trust Management, introduced by Blaze et al. [5], is a unified approach to
specifying and interpreting security policies, credentials, and relationships that
allows direct authorization of security-critical actions. In particular, a trust-
management system combines the notion of specifying security policy with the
mechanism for specifying security credentials. Credentials describe specific del-
egations of trust among public keys; unlike traditional certificates, which bind
keys to names, trust-management credentials bind keys directly to authoriza-
tions to perform specific tasks. Trust-management systems support delegation,
and policy specification and refinement at the different layers of a policy hier-
archy, thus solving to a large degree the consistency and scalabiUty problems
inherent in traditional ACLs. Furthermore, trust-management systems are by
design extensible and can express policies for different types of applications.
Section 2 gives an overview of Trust Management and briefly describes two
tools that we have developed {PolicyMaker and KeyNote). Section 3 presents
some applications of Trust Management. Finally, Section 4 discusses future work
and concludes this paper.

2 Trust Management

This section explains the essence of the trust-management approach, describes

PolicyMaker [5, 7] and KeyNote [4], and ends with a brief discussion of related
trust-management work.

2.1 Basics

A traditional "system-security approach" to the processing of a signed request

for action treats the task as a combination of authentication and access control.
The receiving system first determines who signed the request and then queries
an internal database to decide whether the signer should be granted access to
the resources needed to perform the requested action. We believe that this is the
wrong approach for today's dynamic, internetworked world. In a large, hetero-
geneous, distributed system, there is a huge set of people (and other entities)
who may make requests, as well as a huge set of requests that may be made.
These sets change often and cannot be known in advance. Even if the question
"who signed this request?" could be answered reliably, it would not help in de-
ciding whether or not to take the requested action if the requester is someone
or something from whom the recipient is hearing for the first time.
The right question in a far-flung, rapidly changing network becomes "is the
key that signed this request authorized to take this action?" Because name-
key mappings and pre-computed access-control matrices are inadequate, one
^ For the purposes of this discussion, we consider the traditional capability systems as
ACLs, as they exhibit most of the weaknesses we mention.
 189

needs a more flexible, more "distributed" approach to authorization. The trust-

management approach, initiated by Blaze et al. [5], frames the question as follows:
"Does the set C of credentials prove that the request r complies with the local
security policy P?" Each entity that receives requests must have a policy that
serves as the ultimate source of authority in the local environment. The policy
may directly authorize certain keys to take certain actions, but more typically
it will delegate this responsibility to credential issuers that it trusts to have the
required domain expertise as well as relationships with potential requesters. The
trust-management engine is a separate system component that takes (r, C, P)
as input, outputs a decision about whether compliance with policy has been
proven, and may also output some additional information about how to proceed
if it hasn't.
An essential part of the trust-management approach is the use of a general-
purpose, application-independent algorithm for checking proofs of compliance.
Why is this a good idea? Since any product or service that requires some form
of proof that requested transactions comply with policies could use a special-
purpose algorithm implemented from scratch, what do developers, administra-
tors, and users gain by using a general-purpose compliance checker?
The most important gain is in soundness and reliability of both the defi-
nition and the implementation of "proof of compliance." Developers who set
out to implement a "simple," special-purpose compliance checker (in order to
avoid what they think are the overly "complicated" syntax and semantics of a
universal "meta-policy") may discover that they have underestimated their ap-
plication's need for proof and expressiveness. As they discover the full extent of
their requirements, they may ultimately wind up implementing a system that
is as general and expressive as the "complicated" one they set out to avoid. A
general-purpose compliance checker can be explained, formalized, proven cor-
rect, and implemented in a standard package, and applications that use it can
be assured that the answer returned for any given input (r, C, P) depends only
on the input and not on any implicit policy decisions (or bugs) in the design or
implementation of the compliance checker.
Basic questions that must be answered in the design of a trust-management
engine include:

— How should "proof of compliance" be defined?

— Should policies and credentials be fully or only partially programmable? In
which language or notation should they be expressed?
— How should responsibility be divided between the trust-management engine
and the calling application? For example, which of these two components
should perform the cryptographic signature verification? Should the appli-
cation fetch all credentials needed for the compliance proof before the trust-
management engine is invoked, or may the trust-management engine fetch
additional credentials while it is constructing a proof?

In the rest of this section, we survey several recent and ongoing trust-manage-
ment projects in which different answers to these questions are explored.
 190

2.2 Policy Maker

PolicyMaker was the first example of a "trust-management engine." That is,

it was the first tool for processing signed requests that embodied the "trust-
management" principles articulated in Section 2.1. It addressed the authoriza-
tion problem directly, rather than handling the problem indirectly via authenti-
cation and access control, and it provided an application-independent definition
of "proof of compliance" for matching up requests, credentials, and policies. Pol-
icyMaker was introduced in the original trust-management paper by Blaze et
al. [5], and its compliance-checking algorithm was later fleshed out in [7]. We
give a high-level overview of the design decisions that went into PolicyMaker in
this section and some technical details of the PolicyMaker compliance checker in
Appendix A below. A full description of the system can be found in [5, 7], and
experience using it in several applications is reported in [6, 18, 21].
PolicyMaker credentials and policies are fully programmable; together cre-
dentials and policies are referred to as "assertions." Roughly speaking, assertions
are represented as pairs (/, s), where s is the source of authority, and / is a pro-
gram describing the nature of the authority being granted as well as the party
or parties to whom it is being granted. In a policy assertion, the source is always
the keyword POLICY. For the PolicyMaker trust-management engine to be
able to make a decision about a requested action, the input supplied to it by
the calling application must contain one or more policy assertions; these form
the "trust root," i.e., the ultimate source of authority for the decision about
this request. In a credential assertion, the source is the public key of the issuing
authority. Credentials must be signed by their issuers, and these signatures must
be verified before the credentials can be used.
PolicyMaker assertions can be written in any programming language that
can be "safely" interpreted by a local environment that has to import credentials
from diverse (and possibly untrusted) issuing authorities. A safe version of AWK
was developed for early experimental work on PolicyMaker (see Blaze et al. [5]),
because AWK's pattern-matching constructs are a convenient way to express
authorizations. For a credential assertion issued by a particular authority to be
useful in a proof that a request complies with a policy, the recipient of the request
must have an interpreter for the language in which the assertion is written.
Thus, it would be desirable for assertion writers ultimately to converge on a
small number of assertion languages so that receiving systems have to support
only a small number of interpreters and so that carefully crafted credentials
can be widely used. However, the question of which languages these will be was
left open by the PolicyMaker project. A positive aspect of PolicyMaker's not
insisting on a particular assertion language is that all of that work that has
gone into designing, analyzing, and implementing the PolicyMaker compliance-
checking algorithm will not have to be redone every time an assertion language
is changed or a new language is introduced. The "proof of compliance" and
"assertion-language design" problems are orthogonal in PolicyMaker and can be
worked on independently.
 191

One goal of the PolicyMaker project was to make the trust-management en-
gine minimal and analyzable. Architectural boundaries were drawn so that a
fair amount of responsibility was placed on the calling application rather than
the trust-management engine. In particular, the calling application was made
responsible for all cryptographic verification of signatures on credentials and
requests. One pleasant consequence of this design decision is that the applica-
tion developer's choice of signature scheme(s) can be made independently of his
choice of whether or not to use PolicyMaker for compliance checking. Another
important responsibility that was assigned to the calling application is creden-
tial gathering. The input (r, C, P) supplied to the trust-management module is
treated as a claim that credential set C contains a proof that request r complies
with Pohcy P. The trust-management module is not expected to be able to dis-
cover that C is missing just one credential needed to complete the proof and to
go fetch that credential from e.g., the corporate database, the issuer's web site,
the requester himself, or elsewhere. Later trust-management engines, including
KeyNote [4] and REFEREE [11] divide responsibility between the calling appU-
cation and the trust-management engine differently from the way PolicyMaker
divides it.

The main technical contribution of the PolicyMaker project is a notion of

"proof of compliance" that is fully specified and analyzed. We give an overview
of Pohcy Maker's approach to compliance checking here and some details in Ap-
pendix A; a complete treatment of the comphance checker can be found in [7].
The PolicyMaker runtime system provides an environment in which the pol-
icy and credential assertions fed to it by the calling application can cooperate
to produce a proof that the request comphes with the policy (or can fail to pro-
duce such a proof). Among the requirements for this cooperation are a method
of inter-assertion communication and a method for determining that assertions
have collectively succeeded or failed to produce a proof.
Inter-assertion communication in PolicyMaker is done via a simple, write-
only data structure on which all participating assertions record intermediate re-
sults. Specifically, PolicyMaker initializes the proof process by creating a "black-
board" containing only the request string r and the fact that no assertions have
thus far approved the request or anything else. Then PolicyMaker runs the var-
ious assertions, possibly multiple times each. When assertion {fi,Si) is run, it
reads the contents of the blackboard and then adds to the blackboard one or
more acceptance records {i,Si,Rij). Here Rij is an application-specific action
that source Si approves, based on the partial proof that has been constructed
thus far. Rij may be the input request r, or it may be some related action that
this application uses for inter-assertion communication. Note that the meanings
of the action strings Rij are understood by the application-specific assertion pro-
grams fi, but they are not understood by PolicyMaker. All PolicyMaker does is
run the assertions and maintain the global blackboard, making sure that the as-
sertions do not erase acceptance records previously written by other assertions,
fill up the entire blackboard so that no other assertions can write, or exhibit any
 192

other non-cooperative behavior. PolicyMaker never tries to interpret the action

strings Rij.
A proof of compliance is achieved if, after PoUcyMaker has finished run-
ning assertions, the blackboard contains an acceptance record indicating that
a policy assertion approves the request r. Among the nontrivial decisions that
PolicyMaker must make are (1) in what order assertions should be run, (2) how
many times each assertion should be run, and (3) when an assertion should be
discarded because it is behaving in a non-cooperative fashion. Blaze et al. [7]
provide:

— A mathematically precise formulation of the PolicyMaker compliance-checking

problem.
— Proof that the problem is undecidable in general and is NP-hard even in
certain natural special cases.
— One special case of the problem that is polynomial-time solvable, is useful
in a wide variety of applications, and is implemented in the current version
of PolicyMaker.

Although the most general version of the compliance-checking problem al-

lows assertions to be arbitrary functions, the computationally tractable version
that is analyzed in [7] and implemented in PolicyMaker is guaranteed to be cor-
rect only when all assertions are monotonic. (Basically, if a monotonic assertion
approves action a when given evidence set E, then it will also approve action
a when given an evidence set that contains E; see Appendix A for a formal
definition.) In particular, correctness is guaranteed only for monotonic policy
assertions, and this excludes certain types of policies that are used in practice,
most notably those that make explicit use of "negative credentials" such as revo-
cation lists. Although it is a limitation, the monotonicity requirement has certain
advantages. One of them is that, although the compliance checker may not han-
dle all potentially desirable policies, it is at least analyzable and provably correct
on a well-defined class of policies. Furthermore, the requirements of many non-
monotonic policies can often be achieved by monotonic policies. For example,
the effect of requiring that an entity not occur on a revocation list can also be
achieved by requiring that it present a "certificate of non-revocation"; the choice
between these two approaches involves trade-offs among the (system-wide) costs
of the two kinds of credentials and the benefits of a standard compliance checker
with provable properties. Finally, restriction to monotonic assertions encourages
a conservative, prudent approach to security: In order to perform a potentially
dangerous action, a user must present an adequate set of affirmative credentials;
no potentially dangerous action is allowed "by default," simply because of the
absence of negative credentials.

2.3 KeyNote

KeyNote [4] was designed according to the same principles as PolicyMaker, using
credentials that directly authorize actions instead of dividing the authorization
 193

task into authentication and access control. Two additional design goals for
Key Note were standardization and ease of integration into applications. To ad-
dress these goals, Key Note assigns more responsibility to the trust-management
engine than PolicyMaker does and less to the calling application; for example,
cryptographic signature verification is done by the trust-management engine in
KeyNote and by the application in PolicyMaker. KeyNote also requires that
credentials and policies be written in a specific assertion language, designed to
work smoothly with KeyNote's compliance checker. By fixing a specific and ap-
propriate assertion language, KeyNote goes further than PolicyMaker toward
facilitating efficiency, interoperability, and widespread use of carefully written
credentials and pohcies.
A calling apphcation passes to a KeyNote evaluator a list of credentials, poli-
cies, and requester public keys, and an "Action Environment." This last element
consists of a list of attribute/value pairs, similar in some ways to the Unix^^
shell environment. The action environment is constructed by the calling applica-
tion and contains all information deemed relevant to the request and necessary
for the trust decision. The action-environment attributes and the assignment of
their values must reflect the security requirements of the application accurately.
Identifying the attributes to be included in the action environment is perhaps the
most important task in integrating KeyNote into new applications. The result
of the evaluation is an application-defined string (perhaps with some additional
information) that is passed back to the application. In the simplest case, the
result is something Hke "authorized."
The KeyNote assertion format resembles that of e-mail headers. An example
(with artificially short keys and signatures for readability) is given in Figure 1.

KeyNote-Version: 1
Authorizer: rsa-pkcsl-hex:"1023abcd"
Licensees: dsa-hex:"986512al" I I
rsa-pkcsl-hex:"19abcd02"
Comment: Authorizer delegates read
access t o e i t h e r of the
Licensees
Conditions: ( $ f i l e == "/etc/passwd" &&
$access == "read") ->
{return "ok"}
Signature: rsa-mdS-pkcsl-hex:"fOOf5673"

Fig. 1. Sample KeyNote assertion

As in PolicyMaker, policies and credentials (collectively called assertions)

have the same format. The only difference between policies and credentials is
that a policy (that is, an assertion with the keyword P O L I C Y in the Authorizer
field) is locally trusted (by the compliance-checker) and thus needs no signature.
 194

KeyNote assertions are structured so that the Licensees field specifies ex-
phcitly the principal or principals to which authority is delegated. Syntactically,
the Licensees field is a formula in which the arguments are public keys and the
operations are conjunction, disjunction, and threshold. The semantics of these
expressions are specified in [4].
The programs in KeyNote are encoded in the Conditions field and are essen-
tially tests of the action environment variables. These tests are string compar-
isons, numerical operations and comparisons, and pattern-matching operations.
We chose such a simple language for KeyNote assertions for the following
reasons:

- AWK, one of the first assertion languages used by PohcyMaker, was criticized
as too heavyweight for most relevant applications. Because of AWK's com-
plexity, the footprint of the interpreter is considerable, and this discourages
application developers from integrating it into a trust-management compo-
nent. The KeyNote assertion language is simple and has a minimal-sized
interpreter.
- In languages that permit loops and recursion (including
AWK), it is difficult to enforce resource-usage restrictions, but applications
that run trust-management assertions written by unknown sources often
need to limit their memory- and CPU-usage.
We believe that for out purposes a language without loops, dynamic memory
allocation, and certain other features is sufficiently powerful and expressive.
The KeyNote assertion syntax is restricted so that resource usage is propor-
tional to the program size. Similar concepts have been successfully used in
other contexts [16].
- Assertions should be both understandable by human readers and easy for
a tool to generate from a high-level specification. Moreover, they should be
easy to analyze automatically, so that automatic verification and consistency
checks can done. This is currently an area of active research.
- One of our goals is to use KeyNote as a means of exchanging policy and
distributing access control information otherwise expressed in an applica-
tion-native format. Thus the language should be easy to map to a number
of such formats {e.g., from a KeyNote assertion to packet-filtering rules).
- The language chosen was adequate for KeyNote's evaluation model.

This last point requires explanation.

In PohcyMaker, compliance proofs are constructed via repeated evaluation
of assertions, along with an arbitrated "blackboard" for storage of intermediate
results and inter-assertion communication.
In contrast, KeyNote uses a depth-first search (DPS) algorithm that attempts
(recursively) to satisfy at least one policy assertion. Satisfying an assertion entails
satisfying both the Conditions field and the Licensees key expression. Note that
there is no explicit inter-assertion communication as in PohcyMaker; the accep-
tance records returned by program evaluation are used internally by the KeyNote
evaluator and are never seen directly by other assertions. Because KeyNote's
 195

evaluation model is a subset of PolicyMaker's, the latter's compliance-checking

guarantees are applicable to KeyNote. Whether the more restrictive nature of
KeyNote allows for stronger guarantees to be made is an open question requiring
further research.
Ultimately, for a request to be approved, an assertion graph must be con-
structed between one or more policy assertions and one or more keys that signed
the request. Because of the evaluation model, an assertion located somewhere
in a delegation graph can effectively only refine (or pass on) the authorizations
conferred on it by the previous assertions in the graph. (This principle also holds
for PohcyMaker.) For more details on the evaluation model, see [4].
It should be noted that PolicyMaker's restrictions regarding "negative cre-
dentials" also apply to KeyNote. Certificate revocation lists (CRLs) are not built
into the KeyNote (or the PolicyMaker) system; these however could be provided
at a higher (or lower) level, perhaps even transparently to KeyNote^. The prob-
lem of credential discovery is also not explicitly addressed in KeyNote. We hope
that the fact that KeyNote provides an explicit, descriptive credential format
will facilitate research on both credential discovery and revocation.
Finally, note that KeyNote, like other trust-management engines, does not
directly enforce policy; it only provides advice to the applications that call it.
KeyNote assumes that the application itself is trusted and that the policy asser-
tions are correct. Nothing prevents an application from submitting misleading
assertions to KeyNote or from ignoring KeyNote altogether.

2.4 Related Work on Trust Management

We close with a brief discussion of two general-purpose trust-management sys-

tems that share much of the basic approach initiated by PolicyMaker but depart
from it in some notable ways.
The REFEREE system of Chu et al. [11] is hke PolicyMaker in that it sup-
ports full programmability of assertions (policies and credentials). However, it
differs in several important ways. It allows the trust-management engine, while
evaluating a request, to fetch additional credentials and to perform cryptographic
signature-verification. (Recall that PolicyMaker places the responsibility for both
of these functions on the calling appUcation and insists that they be done before
the evaluation of a request begins.) Furthermore, REFEREE'S notion of "proof
of compliance" is more complex than PolicyMaker's; for example, it allows non-
monotonic policies and credentials. The REFEREE proof system also supports a
more complicated form of inter-assertion communication than PolicyMaker does.
In particular, the REFEREE execution environment allows assertion programs
to call each other as subroutines and to pass different arguments to different sub-
routines, whereas the PolicyMaker execution environment requires each assertion
program to write anything it wants to communicate on a global "blackboard"
that can be seen by all other assertions.

Note that the decision to consult a CRL is (or should be) a matter of local policy.
 196

REFEREE was designed with trust management for web browsing in mind,
but it is a general-purpose language and could be used in other applications.
Some of the design choices in REFEREE were influence by experience (reported
in [6]) with using PolicyMaker for web-page filtering based on PICS labels [26]
and users' viewing policies. It is unclear whether the cost of building and an-
alyzing a more complex trust-management environment such as REFEREE is
justified by the ability to construct more sophisticated proofs of compliance than
those constructible in PolicyMaker. Assessing this tradeoff would require more
experimentation with both systems, as well as a rigorous specification and anal-
ysis of the REFEREE proof system, similar to the one for PolicyMaker given in
[7].
The Simple Public Key Infrastructure (SPKI) project of Ellison et al. [13]
has proposed a standard format for authorization certificates. SPKI shares with
our trust-management approach the belief that certificates can be used directly
for authorization rather than simply for authentication. However, SPKI certifi-
cates are not fully programmable; they are data structures with the following five
fields: "Issuer" (the source of authority), "Subject" (the entity being authorized
to do something), "Delegation" (a boolean value specifying whether or not the
subject is permitted to pass the authorization on to other entities), "Authoriza-
tion" (a specification of the power that the issuer is conferring on the subject),
and "Validity dates." The SPKI certificate format is compatible with the Sim-
ple Distributed Security Infrastructure (SDSI) local-names format proposed by
Rivest and Lampson [19], and Ellison et al. [13] explain how to integrate the
two.
The SPKI documentation [13] states that
The processing of certificates and related objects to yield an authoriza-
tion result is the province of the developer of the application or system.
The processing plan presented here is an example that may be followed,
but its primary purpose is to clarify the semantics of an SPKI certificate
and the way it and various other kinds of certificate might be used to
yield an authorization result.
Thus, strictly speaking, SPKI is not a trust-management engine according to
our use of the term, because compliance checking (referred to above as "process-
ing of certificates and related objects") may be done in an application-dependent
manner. If the processing plan presented in [13] were universally adopted, then
SPKI would be a trust-management engine. The resulting notion of "proof of
compliance" would be considerably more restricted than PolicyMaker's; essen-
tially, proofs would take the form of chains of certificates. On the other hand,
SPKI has a standard way of handling certain types of non-monotonic policies,
because validity periods and simple CRLs are part of the proposal.

3 Applications of Trust Management

In this section, we give a brief overview of experiences with the trust-management
approach in mobile-code security, active networking, nd distributed access-control.
 197

3.1 Active Networks

There has been a great deal of interest in the problem of exposing the abihty
to control of network infrastructure. Much of this interest has been driven by a
the desire to accelerate service creation. Sometimes services can be created using
features of existing systems. One of the most aggressive proposals is the notion
of programmable network infrastructure or "active networking." In an active
network, the operator or user has facilities for directly modifying the operational
semantics of the network itself. Thus, the role of network and endpoint become
far more malleable for the construction of new applications. This is in contrast to
the "service overlay model" as employed, for example, in present-day Internet,
where service introduction at the "edge" of the virtual infrastructure is very easy,
but changes in the infrastructure itself have proven very difficult {e.g., RSVP
[8] and multicasting [12]). A number of active network architectures have been
proposed and are under investigation [1, 28, 16, 27, 9, 25].
A programmable network infrastructure is potentially more vulnerable to
attacks since a portion of the control plane is intentionally exposed, and this
can lead to far more complex threats than exist with an inaccessible control
plane. For example, a denial-of-service attack on the transport plane may also
inhibit access to the control plane. Unauthenticated access to the control plane
can have severe consequences for the security of the whole network.
It is therefore especially necessary for an active network to use a robust and
powerful authorization mechanism. Because of the many interactions between
network nodes and switchlets (pieces of code dynamically loaded on the switches,
or code in packets that is executed on every active node they encounter), a
versatile, scalable, and expressive mechanism is called for.
We have applied KeyNote [4] in one proposed active network security archi-
tecture, in the Secure Active Network Environment (SANE) [2] [3] developed at
the University of Pennsylvania as part of the Switch Ware project [1].
In SANE, the principals involved in the authorization and policy decisions
in the security model are users, programmers and administrators and network
elements. The network elements are presumed to be under physical control of
an administrator. Programmers may not have physical access to the network
element, but may possess considerable access rights to resources present in the
network elements. Users may have access to basic services {e.g., transport), but
also resources that the network elements are willing to export to all users, at an
appropriate level of abstraction. Users may also be allowed to introduce their
own services, or load those written by others. In such a dynamic environment,
KeyNote is used to supply policy and authorization credentials for those compo-
nents of the architecture that enforce resource usage and access control limits.
In particular, KeyNote policies and credentials are used to:

— Authorize principals to load code on active routers. This happens as a result

of an authentication mechanism, further described in [2].
- Intimately related to the previous item, KeyNote sets resource limits {e.g.,
memory allocated, CPU cycles consumed, memory bandwidth) that the ex-
 198

ecution environment'' enforces. The KeyNote evaluator is initially invoked

just before a switchlet begins execution, and is provided with the switchlet
owner's credentials and the node's policies. It then determines whether the
switchlet is allowed to execute on the node (addressing the previous item),
and what the initial resource limits are. If these are exceeded, another call
to the KeyNote system is made, whereupon the decision may be to grant
more resources, kill the switchlet, or raise an exception.
— Fine-grained control of what actions a switchlet may take on the active
node. In our current approach, loaded code may call other loaded or resident
functions (resembling inter-process communication and system call invoca-
tion respectively). Which of these calls are allowed by policy to occur is (or
rather, can optionally be) controlled by KeyNote. For performance reasons,
and since the switchlet language used lends itself to some degree of static
analysis, this authorization step happens at hnk time (when a newly-loaded
switchlet is dynamically linked in the runtime system). An alternative way
of achieving this would be to trap into the KeyNote evaluator every time
an external (to the switchlet) function was called, by generating appropri-
ate function stubs at link time. This could lead to considerable performance
problems, especially for time-critical applications (such as real-time content
processing and distribution). Such functionality, however, is necessary if one
needs to restrict the range of arguments that may be passed to external
functions. This is currently an area of open research.
- KeyNote credentials can be used by "active firewalls" to notify nodes behind
the firewall that a particular piece of active code should (or should not) be
allowed to perform specific tasks [17]. In this context, KeyNote defines trust
relations between active nodes {e.g., "firewall", "protected host", etc)

Trust management for active networks is an area of continuing research.

3.2 Mobile Code Security

Another area of broad recent interest is the security and containment of un-
trusted "mobile" code. That is, executable content or mobile code is received
by a host with a request to a execute it; lacking any automatic mechanisms for
evaluating the security implications of executing such a piece of code, the host
needs to find some other way of determining the trustworthiness of that code.
Failure to properly contain mobile code may result in serious damage or
leakage or information resident on the host. Such damage can be the result of
malicious intent {e.g., industrial or otherwise espionage or vandalism), or un-
intentional {e.g., programming failures or unexpected interactions with other
system components or programs). Other consequences of failing to contain mo-
bile code include denial-of-service attacks (the now familiar phenomenon of Java
and JavaScript applets using all the system memory or CPU cycles, usually

'^ In the case of SANE, the execution environment is a modified and restricted runtime
of the Caml [20] programming language.
 199

because of bugs), or infiltration of a company network through a downloaded

trojan horse or virus.
Of course, these threats existed long before mobile code, such as Java applets,
became popular. Downloading software from a bulletin board service or through
the Internet and then executing it without any form of code review or execution
restrictions has almost exactly the same security implications. The problem,
however, has been made more prominent because of the scale and ease with which
attacks to individuals and hosts can be carried out, especially in environments
where mobile code is run automatically.
As we mentioned in the introduction, various proposals exist that tie code
"trustworthiness" to the existence of a digital signature or other form of authen-
tication. Such mechanisms, however, do not address the underlying problem; a
signature has value only to the extent that the verifier can evaluate the signer's
trustworthiness (which may be a result of poorly-defined, or incorrectly under-
stood, social factors, technical competence, personal knowledge, etc.). In that
respect, signature-based mobile code schemes do not scale. Furthermore, some
programs may be "safe" to run under restricted environments or a small set
of platforms. Simple digital signature schemes as have been proposed cannot
readily accommodate such conditions.
Trust Management has at least two different roles in mobile code security:

- Express trust relations between code-certifying entities, and the conditions

under which their certification has meaning. So, for example, code-certifier
A could state that his certification of safety of a particular piece of code
is predicated by some set of conditions (presumably the conditions under
which the code evaluation took place, or the assumptions made by the code
that may have security implications if violated). Trust management could
also be used to express the local user's (or corporate) policy on executing
mobile code {e.g., "must be signed by the company security officer or must
be accompanied by a PCC proof" [24, 23]).
- Trust-management credentials could be used to describe the minimal set of
capabilities the host environment must grant to enable the code to perform
its tasks. This would then be combined with the relevant local policy® to
restrict the system interface presented to the mobile code. This is similar
in concept to the approach mentioned in Section 3.1. We are also investi-
gating the application of this method in a traditional Unix kernel, in an
environment where language-based protection is not available.

3.3 Access Control Distribution

An indirect, but architecturally very important, benefit of trust management

involves the distribution of traditional access control list (ACL) databases. Some
applications lend themselves naturally to a well-defined ACL-style interface for
describing their security characteristics that do not obviously require the rich
This would happen automatically, in the process of compliance checking.
 200

expressiveness of a trust-management approach. For example, it may be natural

to describe who is allowed to log in to a computer according to a standard list of
login accounts. Even when an ACL provides sufficient expressiveness, however,
it is often architecturally beneficial to implement the ACL on top of a trust-
management system.
Architectures based on a trust-management system can be easily extended
if, in the future, it becomes necessary to base access decisions on more complex
rules than are captured by an ACL. For example, it is natural and easy to
add "time-of-day" restrictions as part of a policy or individual user credentials,
even if the need for such restrictions had not been anticipated at the time the
system was first designed. Of course, extensibility is a general benefit of trust
management that is not limited to ACL applications.
More importantly, a trust-management system decouples the specification of
access control policies from the mechanism used to distribute and implement
them. For example, consider a traditional access control list maintained as a
simple database in a computer system. A trust-management tool could convert
ACL entries into credentials automatically, with a general credential (certifi-
cate) distribution mechanism used to ensure that updated credentials are sent
to the appropriate place. Distributed, reliable access control can be a very dif-
ficult problem when implemented by itself, and is greatly simplified using a
credential-based trust-management scheme as a back end to a non-distributed
access control database. Even if the expressive power is limited to simple ACLs,
trust management allows the use of an existing credential distribution scheme
to solve the most difficult aspects of the problem.

4 Conclusion and F u t u r e Work

In the time since "trust management" first appeared in the literature in [5], the
concept has gained broad acceptance in the security research community. Trust
management has a number of important advantages over traditional approaches
such as distributed ACLs, hardcoded security pohcies, and global identity cer-
tificates. A trust-management system provides direct authorization of security-
critical actions and decouples the problem of specifying policy and authorization
from that of distributing credentials.
Our work on trust management has focused on designing languages and com-
pliance checkers, identifying applications, and building practical toolkits. There
are important areas that we have not yet addressed. Foremost among those is
automated credential discovery; in our current systems, it is the responsibility of
the requester to submit all necessary credentials, under the assumption that he
holds all credentials relevant to him. Even then, however, intermediate creden-
tials that form the necessary connections between the verifier's policy and the
requester's credentials must be acquired and submitted. The various solutions
to this problem range from "leave it to the application" to using distributed
databases and lookup mechanisms for credential discovery. A solution along the
 201

lines of a distributed d a t a b a s e can also assist large organizations in auditing

their security and access policies.
Another area of future work is providing functionality similar t o t h a t of
certificate revocation lists (CRLs), keeping in mind the constraints a b o u t "nega-
tive credentials" mentioned in Section 2.2. While adoption and deployment of a
scheme similar to t h a t of X.509 is fairly straightforward, we are still investigating
the security and operational implications.
Finally, we are examining higher-level policy languages t h a t are even more
human-understandable and capable of higher levels of abstraction. Such high-
level policy would be combined with network- and application-specific infor-
mation and compiled into a set to trust-management credentials. Similarly, a
tool for translating trust-management credentials into apphcation-native forms
would give us all the advantages of t r u s t management (delegation, formal proof
of compliance, etc.) while requiring minimal changes to applications.

References

[1] D. S. Alexander, W. A. Arbaugh, M. Hicks, P. Kakkar, A. D. Keromytis, J. T.

Moore, C. A. Gunter, S. M. Nettles, and J. M. Smith. The SwitchWare Ac-
tive Network Architecture. IEEE Network Magazine, special issue on Active and
Programmable Networks, 12(3):29-36, 1998.
[2] D. S. Alexander, W. A. Arbaugh, A. D. Keromytis, and J. M. Smith. A Secure
Active Network Environment Architecture: Realization in SwitchWare. IEEE
Network Magazine, special issue on Active and Programmable Networks, 12(3):37-
45, 1998.
[3] D. S. Alexander, W. A. Arbaugh, A. D. Keromytis, and J. M. Smith. Security in
active networks. In Jan Vitek and Christian Jensen, editors. Secure Internet Pro-
gramming, Lecture Notes in Computer Science. Springer-Verlag Inc., New York,
NY, USA, 1999.
[4] M. Blaze, J. Feigenbaum, J. loannidis, and A. Keromytis.
The KeyNote Trust-Management System. Work in Progress,
http://www.CIS.upenn.edu/'angelos/keynote.html, June 1998.
[5] M. Blaze, J. Feigenbaum, and J. Lacy. Decentralized Trust Management. In Proc.
of the 17th Symposium on Security and Privacy, pages 164-173. IEEE Computer
Society Press, Los Alamitos, 1996.
[6] M. Blaze, J. Feigenbaum, P. Resnick, and M. Strauss. Managing Trust in an
Information Labeling System. In European Transactions on Telecommunications,
8, pages 491-501, 1997.
[7] M. Blaze, J. Feigenbaum, and M. Strauss. Comphance Checking in the Policy-
Maker Trust-Management System. In Proc. of the Financial Cryptography '98,
Lecture Notes in Computer Science, vol. 1465, pages 254-274. Springer, Berlin,
1998.
[8] R. Braden, L. Zhang, S. Berson, S. Herzog, and S. Jamin. Resource ReSerVation
Protocol (RSVP) - Version 1 Functional Specification. Internet RFC 2208, 1997.
[9] M. Calderon, M. Sedano, A. Azcorra, and C. Alonso. The Support of Active
Networks for Fuzzy-Tolerant Multicast Applications. IEEE Network Magazine,
special issue on Active and Programmable Networks, 12(3):20-28, 1998.
 202

[10] J. Chinitz and S. Sonnenberg. A Transparent Security Framework For T C P / I P

and Legacy Applications. Technical report, Intellisoft Corp., August 1996.
[11] Y.-H. Chu, J. Feigenbaum, B. LaMacchia, P. Resnick, and M. Strauss. REFEREE:
Trust Management for Web Applications. In World Wide Web Journal, 2, pages
127-139, 1997.
[12] S. E. Deering. Host extensions for IP multicasting. Internet RFC 1112, 1989.
[13] C. M. Ellison, B. Frantz, R. Rivest, B. M. Thomas, and
T. Ylonen. Simple Pubhc Key Certificate. Work in Progress,
http://www.pobox.com/"cme/html/spki.html, April 1997.
[14] S. Even, A. Salman, and Y. Yacobi. The Complexity of Promise Problems with
Applications to Public-Key Cryptography. Information and Control, 61:159-174,
1984.
[15] James Gosling, Bill Joy, and Guy Steele. The Java Language Specification. Ad-
dison Wesley, Reading, 1996.
[16] M. Hicks, P. Kakkar, J. T. Moore, C. A. Gunter, and S. Nettles. PLAN: A
Programming Language for Active Networks. Technical Report MS-CIS-98-25,
Department of Computer and Information Science, University of Pennsylvania,
February 1998.
[17] Angelos D. Keromytis, Matt Blaze, John loannidis, and Jonathan M. Smith. Fire-
walls in Active Networks. Technical Report MS-CIS-98-03, University of Penn-
sylvania, February 1998.
[18] J. Lacy, J. Snyder, and D. Maher. Music on the Internet and the Intellectual Prop-
erty Protection Problem. In Proc. of the International Symposium on Industrial
Electronics, pages SS77-83. IEEE Press, 1997.
[19] B. Lampson and R. Rivest. Cryptography and Information Security Group Re-
search Project: A Simple Distributed Security Infrastructure. Technical report,
MIT, 1997.
[20] Xavier Leroy. The Caml Special Light System (Release 1.10).
http://pauillac.inria.fr/ocaml.
[21] R. Levien, L. McCarthy, and M. Blaze. Transparent Internet E-mail Security.
http://www.cs.umass.edu/~lmccarth/crypto/papers/email.ps.
[22] S. P. Miller, B, C. Neuman, J. I. Schiller, and J. H. Saltzer. Kerberos authenti-
cation and authorization system. Technical report, MIT, December 1987.
[23] George C. Necula. Proof-Carrying Code. In Proceedings of the 24th Annual
ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages
(POPL), pages 106-119. ACM Press, New York, January 1997.
[24] George C. Necula and Peter Lee. Safe Kernel Extensions Without Run-Time
Checking. In Second Symposium on Operating System Design and Implementation
(OSDI), pages 229-243. Usenix, Seattle, 1996.
[25] C. Partridge and A. Jackson. Smart Packets. Technical report, BBN, 1996.
http://www.net-tech.bbn.com-/smtpkts/smtpkts-index.html.
[26] P. Resnick and J. Miller. PICS: Internet Access Controls Without Censorship.
Communications of the ACM, pages 87-93, October 1996.
[27] D. Wetherall, U. Legedza, and J. Guttag. Introducing New Internet Services: Why
and How. IEEE Network Magazine, special issue on Active and Programmable
Networks, 12(3):12-19, 1998.
[28] David J. Wetherall, John Guttag, and David L. Tennenhouse. Ants: A toolkit
for building and dynamically deploying network protocols. In IEEE OpenArch
Proceedings. IEEE Computer Society Press, Los Alamitos, April 1998.
 203

A PolicyMaker Compliance Proofs

We now give some technical details about the PolicyMaker compliance checker.
This discussion is largely excerpted (with the authors' permission) from the
paper of Blaze et al. [7].
The general problem we are concerned with is Proof of Compliance (POC).
The question is whether a request r complies with a policy. The policy is simply
a function /o encoded in some well understood programming system or language
and labeled by the keyword POLICY. In addition to the request and the policy,
a POC instance contains a set of credentials, also general functions, each labeled
by its source. Policies and credentials are collectively referred to as assertions.
Credentials are issued by sources. Formally, a credential is a pair (/t,Sj)
of function fi and source-ID Si, which is just a string over some appropriate
alphabet. Important examples of source-IDs include public keys of credential
issuers, URLs, names of people, and names of companies. With the exception of
the keyword POLICY, the interpretation of source-IDs is part of the apphcation-
specific semantics of an assertion, and it is not the job of the compliance checker.
From the compliance checker's point of view, the source-IDs are just strings, and
the assertions encode a set of (possibly indirect and possibly conditional) trust
relationships among the issuing sources. Associating each assertion with the
correct source-ID is the responsibility of the calling application, as explained in
Section 2.2.
The request r is a string encoding an action for which the calling applica-
tion seeks a proof of compliance. In the course of deciding whether the creden-
tials (/i,si), • • •, {fn-i,Sn-i) constitute a proof that r complies with the policy
(/o, POLICY), the comphance checker's domain of discourse may need to include
other action strings. For example, if POLICY requires that r be approved by
credential issuers Si and s-z, the credentials (/i, Si) and (/2, S2) may want a way
to say that they approve r conditionally, where the condition is that the other
credential also approve it. A convenient way to formalize this is to use strings
R, Ri, and R2 over some finite alphabet S. The string R corresponds to the
requested action r. The strings Ri and R2 encode "conditional" versions of R
that might be approved by si and S2 as intermediate results of the compliance-
checking procedure.
More generally, for each request r and each assertion {fi,Si), there is a set
{Rij} of action strings that might arise in a compliance check. By convention,
there is a distinguished string R that corresponds to the input request r. The
range of assertion {fi, Si) is made up of acceptance records of the form (i, Sj, Rij),
the meaning of which is that, based on the information at its disposal, assertion
number i, issued by source Si, approves action Rij. A set of acceptance records is
referred to as an acceptance set. It is by maintaining acceptance sets and making
them available to assertions that the PolicyMaker compliance checker manages
"inter-assertion communication," giving assertions the chance to make decisions
based on conditional decisions by other assertions. The compliance checker will
start with initial acceptance set {{A, A, R)}, in which the one acceptance record
means that the action string for which approval is sought is R and that no as-
 204

sertions have yet signed off on it (or anything else). The checker will run the
assertions (/o,POLICY), ( / i , s i ) , . . . , (/„_i, s„_i) that it has received as input,
not necessarily in that order and not necessarily once each, and see which ac-
ceptance records are produced. Ultimately, the compliance checker approves the
request r if the acceptance record (0, POLICY, R), which means "policy approves
the initial action string," is produced.
Thus, abstractly, an assertion is a mapping from acceptance sets to accep-
tance sets. Assertion {fi,Si) looks at an acceptance set A encoding the actions
that have been approved so far and the numbers and sources of the assertions
that approved them. Based on this information about what the sources it trusts
have approved, {fi,Si) outputs another acceptance set A'.
The following concrete examples show why PolicyMaker assertions are al-
lowed to approve multiple action strings for each possible request. That is, for
a given input request r, why do assertions need to do anything except say "I
approve r" or refuse to say it?
First, consider the following "co-signing required" assertion (/o,POLICY):
"All expenditures of $500 or more require approval by A and B." Suppose that
A's policy is to approve such expenditures if and only if B approves them and
that B's is to approve them if and only if A approves them. Our acceptance
record structure makes such approvals straightforward. The credential (/i,A),
can produce acceptance records of the form {1,A,R) and {1,A,RB), where R
corresponds to the input request r; the meaning of the second is "I will approve
R if and only if B approves it." Similarly, the credential (/2,B), can produce
records of the form {2,B,R) and {2,B,RA)- On input {{A,A,R)}, the sequence
of acceptance records (1, A, fis), {2,B,RA), (l,A,i?), (2,3, R), (0, POLICY, i?)
would be produced if the assertions were run in the order (/i. A), (/2,B), (/i, A),
(/2,B), (/o, POLICY), and the request r would be approved. If assertions could
only produce binary approve/disapprove decisions, no transactions would ever be
approved, unless the trust management system had some way of understanding
the semantics of the assertions and knowing that it had to ask A's and B's
credentials explicitly for a conditional approval. This would violate the goal of
having a general-purpose, trust management system that processes requests and
assertions whose semantics are only understood by the calling applications and
that vary widely from application to application.
Second, consider the issue of "delegation depth." A very natural construction
to use in assertion (/o, POLICY) is "I delegate authority to A. Furthermore, I
allow A to choose the parties to whom he will re-delegate the authority I've
delegated to him. For any party B involved in the approval of a request, there
must be a delegation chain of length at most two from me to B." Various "domain
experts" Bi, ..., Bt could issue credentials ( / i , B i ) , ..., {ft,Bt) that directly
approve actions in their areas of expertise by producing acceptance records of
the form {i,Bi,Rl). An assertion {gj,Sj) that sees such a record and explicitly
trusts Bi could produce an acceptajice record of the form {j, Sj,R\), the meaning
of which is that "Bi approved R^ directly, I trust Bi directly, and so I also
approve i ? \ " More generally, if an assertion (<;;, s;) trusts s/t directly and sees an
 205

acceptance record of the form (fe,Sfc,i?J;), it can produce the acceptance record
{l,si,R'^^^). The assertion (/o,POLICY) given above would approve an action
W if and only if it were run on an acceptance set that contained a record of
the form {k,A,R\), for some k. Note that (/o,POLICY) need not know which
credential {fi,Bi) directly approved i?' by producing {i,Bi,R]^). All it needs to
know is that it trusts A and that A trusts some Bi whose credential produced
such a record.
The most general version of the compliance-checking problem is:
Proof of Compliance (POC):
Input: A request r and a set {(/o,POLICY), (/i,.si), ..., ( / „ - i , s „ - i ) } of
assertions.
Question : Is there a finite sequence i i , i 2 , . . . , it of indices such that each ij is
in { 0 , 1 , . . . ,n — 1}, but the ij's are not necessarily distinct and not necessarily
exhaustive o f { 0 , l , . . . , n — 1} and such that

(0, POLICY, ii)e(/„, Si, )o...o(/,,,s,j({(^^ii:)}),

where R is the action string that corresponds to the request r?
This most general version of the problem is clearly undecidable. A compliance
checker cannot even decide whether an arbitrary assertion (/j,Sj) halts when
given an arbitrary acceptance set as input, much less whether some sequence
containing [fi, Si) produces the desired output.
When we say that "{(/o,POLICY), ( / i , s i ) , . . . , (/„_i, s„_i)} contains a
proof that r complies with POLICY," we mean that (r, {(/o, POLICY), (/i, Si),
..., (/n-i,s„_i)}) is a yes-instance of this unconstrained, most general form
of POC. If F is a (possibly proper) subset of {(/o,POLICY), ( / i , s i ) , ...,
(/„_i,Sn_i)} that contains all of the assertions that actually appear in the
sequence {fi^, SjJ o • • • o [f^^, SiJ, then we say that "F contains a proof that r
complies with POLICY."
Restricted versions of POC are obtained by adding various pieces of infor-
mation to the problem instances. Specifically, consider augmenting the instance
(r, {(/o, POLICY), ( / i , s i ) , ..., (/„_i,s„_i)}) in one or more of the following
ways:
Global runtime bound: An instance may contain an integer d such that a
sequence of assertions (/jj, Si^), ..., {fi^, Si,) is only considered a valid proof
that r comphes with POLICY if the total amount of time that the compliance
checker needs to compute (/i,,Si,) o • • • o {fi^,Si^){{{A,A,R)}) is 0(N''). Here
N is the length of the original problem instance, i.e., the number of bits needed
to encode r, (/Q,POLICY), ..., (/«_i,.s,i_i), and d in some standard fashion.
Local runtime bound: An instance may contain an integer c such that (/ij,
Sii), • • •, (/it, Si,) is only considered a vahd proof that r complies with POLICY if
each {fi-, Si-) runs in time 0{N^). Here A'' is the length of the actual acceptance
set that is input to {fi-, Sj^.) when it is run by the compliance checker. Note
that the length of the input fed to an individual assertion {fij, Si.) in the course
of checking a proof may be considerably bigger than the length of the original
 206

problem instance (r, {(/o,POLICY), ( / i , s i ) , . . . , (/„_i,s„_i)},c), because the

running of assertions (/ii,SiJ, ..., (/ij_i,Si,_i) may have caused the creation
of many new acceptance records.
Bounded number of assertions in a proof: An instance may contain an
integer I such that (/iuSi,), • • •, {fit,Si,) is only considered a valid proof if i < I.
Bounded output set: An instance may contain integers m and s such that an
assertion (/,, Si) can only be part of a valid proof that r complies with POLICY if
there is a set Oi = {Ru,..., B.i,n} of m action strings, such that {fi,Si){A) C Oj
for any input set A, and the maximum size of an acceptance record {i, Si, Rij) is s.
Intuitively, for any user-supplied request r, the meaningful "domain of discourse"
for assertion {ft, Si) is of size at most m — there are at most m actions that it
would make sense for {fi,Si) to sign off on, no matter what the other assertions
in the instance say about r.
Monotonicity: Important variants of POC are obtained by restricting atten-
tion to instances in which the assertions have the following property: (/i,Sj) is
monotonic if, for all acceptance sets A and B, A C B => {fi, Si){A) C (/j, Si)(B).
Thus, if ifijSi) approves action B,ij when given a certain set of "evidence" that
Rij is ok, it will also approve Rij when given a superset of that evidence — it
does not have a notion of "negative evidence."
Any of the parameters I, m, and s that are present in a particular instance
should be written in unary so that they play an analogous role to n (the number
of assertions) in the calculation of the total size of the instance. The parameters d
and c are exponents in a runtime bound and hence can be written in binary. Any
subset of the parameters d, c, /, m, and s may be present in a POC instance,
and each subset defines a POC variant, some of which are more natural and
interesting than others. Including a global runtime bound d obviously makes the
POC problem decidable, as does including parameters c and /.
In stating and proving results about the complexity of POC, we use the notion
of a promise problem [14]. In a standard decision problem, a language L is defined
by a predicate R in that x G L <^ R(x). In a promise problem, there are two
predicates, the promise Q and the property R. A machine M solves the promise
problem iQ,R) if, for all inputs x for which the promise holds, the machine M
halts and accepts x if and only if the property holds. Formally, Vx[Q(x) ^ [M
halts on x and M{x) accepts <=> .R(.T)]]. Note that M's behavior is unconstrained
on inputs that do not satisfy the promise, and each set of choices for the behavior
of M on these inputs determines a different solution. Thus predicates Q and R
define a family of languages, namely all L such that L = L{M) for some M that
solves {Q,R). A promise problem is NP-hard if it has at least one solution and
all of its solutions are NP-hard.
The following natural variants of POC are NP-hard. Refer to Blaze et al. [7]
for the NP-hardness proofs.

Locally Bounded Proof of Compliance (LBPOC):

Input : A request r, a set {(/o, POLICY), ( / i , 6 i ) , . . . , ( / n - i , s „ _ i ) } of assertions,
and integers c, /,TO,and s.
 207

Promise : Each (/i,Sj) runs in time 0{N''). On any input set that contains
{A, A, R), where R is the action string corresponding to request r, for each (/j, Sj)
there is a set Oi of at most m action strings such that {fi,Si) only produces
output from Oi, and s is the maximum size of an acceptance record {i,Si, Rij),
where Rij € Oi.
Question : Is there a sequence ii,...,it of indices such that

1. Each ij is in { 0 , 1 , . . . , n — 1}, but the ij need not be distinct or collectively

exhaustive of { 0 , 1 , . . . , n — 1},
2. t <l, and
3. {0,FOUCY,R)e{h,,Si,)o---o{f,^,s,^){{{A,A,R)})l

Globally Bounded Proof of Compliance (GBPOC):

Input: A request r, a set {(/o, POLICY), ( / i , S i ) , . . . , ( / „ _ i , s „ _ i ) ) of assertions,
and an integer d.
Question : Is there a sequence ii,... ,if of indices such that

1. Each ij is in { 0 , 1 , . . . , n — 1}, but the ij need not be distinct or collectively

exhaustive of { 0 , 1 , . . . , n — 1},
2. (0, POLICY, R) e (/,,, Sijo• • • o(/i,, SiJ({(yl, A, R)}), where R is the action
string corresponding to request r, and
3. The computation of {fi,,Si,)o- • • o{fi^,Sij){{{A, A,R)}) runs in (total) time
OiN")?

Monotonic Proof of Compliance (MPOC):

Input : A request r, a set {(/o, POLICY), ( / i , s i ) , . . . , (/n-i,Sn-i)} of assertions,
and integers I and c.
Promise : Each assertion {Ji, Si) is monotonic and runs in time 0 ( i V ) .
Question : Is there a sequence ii,... ,it of indices such that

1. Each ij is in { 0 , 1 , . . . , n — 1}, but the ij need not be distinct or collectively

exhaustive of {0,1,..., n — 1},
2. t<l, and
3. (0,POLICY, R) e (/,,, Sj,)o• • • o(/,j,Si,){{{A, A,R)}), where R is the action
string corresponding to request r?

Each version of POC can be defined using "agglomeration" [f-^, s-^) * (/i, si)
instead of composition (/2,S2) ° ( / i , s i ) . The result of applying the sequence
of assertions (/ii,SiJ, ..., (/i, ,Si,) agglomeratively to an acceptance set 5o is
defined inductively as follows: Si = (/ii,Sij)(5'o) U So and, for 2 < j < t,
Sj = (fi.,Si.){Sj-i) U Sj-i. Thus, for any acceptance set A, A C {fit,SiJ *
• • • * (/jj, Sjj) (A). The agglomerative versions of the decision problems are iden-
tical to the versions already given, except that the acceptance condition is
"(0,POLICY,i?) e {fi,,Si,)*---*{fi,,Si,){{{A,A,R)})r' We refer to "agglom-
erative POC," "agglomerative MPOC," etc., when we mean the version defined
in terms of * instead of o.
 208

A trust management system that defines "proof of compliance" in terms

of agglomeration makes it impossible for an assertion to "undo" an approval
that it or any other assertion has already given to an action string during the
course of constructing a proof. Informally, it forces assertions to construct proofs
by communicating on a "write-only blackboard." This definition of proof makes
sense if it is important for the trust management system to guard against a rogue
credential-issuer's ability to thwart legitimate proofs. Note that the question
of whether the compliance checker combines assertions using agglomeration or
composition is separate from the question of whether the assertions themselves
are monotonic.
The agglomerative versions of GBPOC, LBPOC, and MPOC are also NP-
hard; the NP-hardness proofs are given in [7] and are simply minor variations
on the NP-hardness proofs for the composition versions.
Finally, we present the compliance-checking algorithm that is used in the
current version of the Policy Maker trust management system. The promise that
defines this special case includes some conditions that we have already discussed,
namely monotonicity and bounds on the runtime of assertions and on the total
size of acceptance sets that assertions can produce. It also includes "authentic-
ity," something that can be ignored when proving hardness results. An authentic
assertion {fi, Si) only produces acceptance records of the form {i, Si,Rij), i.e., it
does not "impersonate" another assertion by producing an acceptance record of
the form («', s^- ,Riij).
PolicyMaker constructs proofs in an agglomerative fashion, and hence we
use * in the following problem statement. This variant of POC could be defined
using o as well, but the algorithm given below would not work for the o version.
Locally Bounded, Monotonic, and Authentic Proof of Compliance
(LBMAPOC):
Input : A request r, a set {(/o, POLICY), (/i,Si),. •., ( / n - i , s „ _ i ) } of assertions,
and integers c, m, and s.
Promise : Each assertion {fi, Si) is monotonic, authentic, and runs in time 0{N'^).
On any input set that contains {A, A, R), where R is the action string correspond-
ing to request r, for each {fi,Si) there is a set O, of at most m action strings,
such that {fi,Si) only produces output from O,, and s is the maximum size of
an acceptance record {i,Si,Rij), such that Rij E Oi.
Question : Is there a sequence ii,...,«« of indices such that each ij is in { 0 , 1 , . . . ,
n—1}, but the ij need not be distinct or collectively exhaustive of { 0 , 1 , . . . , n—1},
and (0, POLICY, i?) £ {h,~,su) * • • • * {fi„Si,){{{A,A,R)]).
The current algorithm is called CCAi, for "compliance-checking algorithm,
version 1," to allow for the evolution of PolicyMaker, and for improved algorithms
CCA,, i > 1.
Assertion {fi, Si) is called "ill-formed" if it violates the promise. If CCAi
discovers in the course of simulating it that {fi,Si) is ill-formed, CCAi ignores
it for the remainder of the computation. Note that an assertion {fi,Si) may
be undetectably ill-formed; for example, there may be sets A C B such that
{fi,Si){A) 2 {fi,Si){B), but such that A and B do not arise in this run of the
 209

compliance checker. The CCAi algorithm checks for violations of the promise
every time it simulates an assertion. The pseudocode for these checks is omitted
from the statement of CCAi given here, because it would not illustrate the basic
structure of the algorithm; the predicate IllFormedi) is included in the main
loop to indicate that the checks are done for each simulation.

F i g . 2 . P s e u d o c o d e for Algorithm C C A i

CCAi(r. { ( f o , POLICY), (f i , s i ) , . . . . (f„-i , s„„i)}, c , m, s ) :

{
S <- {U, A, R)}
I ^ {}
For j <— 1 t o nrn
{
For i *r- n - 1 t o 0
{
If ( f i , S i ) ^ I . Then S ' <- (f, , s , : ) ( S )
If I l l F o r m e d ( ( f , ; , S i ) ) , Then I -(- I U { ( f , , S i ) } ,
E l s e S •(- S U S'
}
}
If ( 0 , POLICY, R) 6 S, Then O u t p u t ( A c c e p t ) ,
Else Output(Reject)
}

Note that CCAi does mn iterations of the sequence ( / „ - i , 5„_i), ..., (/i, si),
(/o, POLICY), for a total of mri^ assertion-simulations. Recall that a set F =
{(/ji>Sji).--->(/i,:Si,)} C {(/o, POLICY), ..., (/„_i, s„_i)} "contains a proof
that r complies with POLICY" if there is some sequence fci,..., A;„ of the indices
ill • • • litj not necessarily distinct and not necessarily exhaustive oi jx,... ,jt,
such that (0,POLICY, i ? ) e (/it„,SfcJ * • • • * (/fc.,SfcJ({(yl,yl,i?)}).
The following formal claim about this algorithm is proven in [7].
Theorem 1. Let (r, {(/o, POLICY), (/i, s i ) , . . . , (/„_i, s„_i)}, c, m, s) be an
(agglomerative) LBMAPOC instance.
(1) Suppose that F C {(/o, POLICY), ( / i , s i ) , . . . , (/„_i,s„_i)} contains a
proof that r complies with POLICY and that every {fi,Si) E F satisfies the
promise of LBMAPOC. Then CCAi accepts (r, {(/o, POLICY), (/i, si), . . . ,
(/n-l, Sn-l) },C,m,s).
(2) If {(/o, POLICY), (/i, s i ) , . •., (/n-i, Sn-i)} does not contain aproof that
r complies with POLICY, then CCAi rejects (r, {(/o, POLICY), (/i, si), . . . ,
[fn-l, Sn-l) ],C,m,s).
 210

(3) CCAi runs in time 0{mn'-^{nmsy).

Note that cases (1) and (2) do not cover all possible inputs to CCAi. There
may be a subset F of the input assertions that does contain a proof that r
complies with POLICY but that contains one or more ill-formed assertions. If
CCAi does not detect that any of these assertions is ill-formed, because their
ill-formedness is only exhibited on acceptance sets that do not occur in this com-
putation, then CCAi will accept the input. If it does detect ill-formedness, then,
as specified here, CCAi may or may not accept the input, perhaps depending
on whether the record (0, POLICY, B.) has already been produced at the time of
detection. CCAi could be modified so that it restarts every time ill-formedness
is detected, after discarding the ill-formed assertion so that it is not used in the
new computation. It is not clear whether this modification would be worth the
performance penalty. The point is simply that CCAi offers no guarantees about
what it does when it is fed a policy that trusts, directly or indirectly, a source
of ill-formed assertions, except that it will terminate in time 0{mn^{nmsy). It
is the responsibility of the poHcy author to know which sources to trust and to
modify the policy if some trusted sources are discovered to be issuing ill-formed
assertions.
Finally, note that 0{mn^ {nmsY) is a pessimistic upper bound on the running
time of the compliance checker. It is straightforward to check (each time an
assertion {fi,Si) is run, or at some other regular interval) whether the acceptance
record (0, POLICY, il) has been produced and to "stop early" if it has. Thus,
for many requests R that do comply with policy, the algorithm CCAi will find
compliance proofs in time less than 0{jnn^{nmsy).
 Distributed Access-Rights Management with
Delegation Certificates

Tuomas A u r a

A b s t r a c t . New key-oriented discretionary access control systems are

based on delegation of access rights with public-key certificates. This pa-
per explains the basic idea of delegation certificates in abstract terms and
discusses their advantages and limitations. We emphasize decentraliza-
tion of authority and operations. The discussion is based mostly on the
SPKI certificates but we avoid touching implementation details. We also
describe how threshold and conditional certificates can add flexibility
to the system. Examples are given of access control between intelligent
networks services.

1 Introduction
New distributed discretionary access control mechanisms such as S P K I [14] and
PolicyMaker [8,9] aim for decentralization of authority and management oper-
ations. They do not rely on a trusted computing base (TCB) like traditional
distributed access control [22]. Instead, the participants are assumed to b e un-
trusted the way computers on open networks (e.g. Internet) are in reality.
The decentralization, however, does not mean sMding back t o anarchy such
as the P G P web of trust [26]. T h e new systems offer ways of building local rela-
tions and setting up local authorities t h a t arise from the personal and business
connections of the participants. The access control mechanisms do not m a n d a t e
any hierarchical or fixed domain structure like, for example, Kerberos [18] a n d
DSSA [16]. All entities are equally entitled to distribute rights t o the services in
their control and to act as an authority for those who depend on t h e m for t h e
services.
The main mechanism used in the new access control systems is delegation
of access rights with signed certificates. T h e signing is done with public-key
cryptography. With a certificate, one cryptographic key delegates some of its
authority to another key. T h e certificates can form a comphcated network t h a t
reflects the underlying relations between the owners of the private signature
keys.
By taking the cryptographic keys as their principal entities, the systems avoid
dependence on trusted name and key services such as the X.500 directory [12].
If any names are used, they are not global distinguished names but relative t o
the users [1,24].
This paper explains the principles behind the delegation certificates in an
abstract setting without exposing the reader to implementation details. T h e dis-
cussion is based primarily on the S P K I draft s t a n d a r d although we will not touch
 212

certificate formats. We stress distribution, scalability and locality of policy deci-

sions. Application examples are given from the Calypso service architecture for
intelligent networks [19]. The emphasis is this paper is on practical issues. More
theoretical treatments of distributed access control can be found in [2,4]. Some
support from implementations is in [3,20]. Code libraries and products using
SPKI are expected to appear when the standardization work nears completion.
Sec. 2 introduces delegation certificates an explains how they aid decen-
tralization. Sec. 3 describes two enhancements, threshold schemes and validity
conditions, that add flexibility to the basic certificates. The limitations of the
certificate-based approach and ways to overcome them are the topic of Sec. 4.
Sec. 5 concludes the paper.

2 Access Control with Delegation Certificates

Sec. 2.1 introduces delegation certificates and some basic concepts. In Sec. 2.2, we
describe first how access rights are distributed and verified with certificates. Then
we consider redelegation and extend the verification procedure to cover chains
of delegation certificates. Sec. 2.3 explains why certificates are an appealing
alternative for other methods of distributed access control.

2.1 Basic Delegation Certificates

Delegation certificate — authorization

subject- v'Key2)-tfiay access service S""

— — — _____^__^_ -validity period
during the period'^ T1-T2.;-*-
Signed:'^Key1_)

issuer

Fig. 1. With a delegation certificate, the issuer shares authority with the subject.

A delegation certificate (Fig. 1) is a signed message with which an entity

grants access rights that it has to another entity. We are interested in systems
where the certificates are signed with public-key cryptography and the entities
granting and receiving access rights are, with some exceptions, cryptographic
keys. A certificate has the meaning:

5/c (During the validity period Ti — T2, if I have any of the rights R,
I give them also to K'.)

{Ski- • •) denotes a signed message that includes both the signature and the
original message.) The key that signed the certificate (K) is the issuer and
 213

the key to whom the rights are given (K') is the subject of the certificate, and
the rights R given by the certificate are the authorization (following the SPKI
terminology). With the certificate, the issuer delegates the rights to the subject.
All delegation certificates have a validity period (Ti — T2) specified on them.
When the certificate expires, the subject loses the rights that the certificate may
have given to it. Together with the authorization field, this parameter is used
for regulating the amount of trust the issuer places on the subject. Extremely
short validity periods are used to force on-line connections to the issuer. (For
simplicity, we often omit the validity period in the text below.)
The authorization is usually the right to use certain services. Sometimes, it
can be an attribute that the subject uses as a credential to acquire access rights.
Such attributes can be interpreted as abstract rights that may not directly entitle
the subject to any services but help in acquiring such rights. The syntax of the
authorization is application dependent and each apphcation must provide its
own rules for comparing and combining authorizations. (See [14] Sec. Examples
for some typical authorizations.)
Some characteristics of delegation certificates are that any key can issue
certificates, a key may delegate rights that it does not yet have but hopes to
obtain, and the issuer itself does not lose the rights it gives to the subject. In
the following, we will discuss these and other properties of delegation in detail.
Our view of the world is key-oriented. The entities possessing, delegating
and receiving access rights are cryptographic key pairs. The public key is used
to identify the key pair and to verify signatures. The private key can sign mes-
sages. It is held secret by some physical entity that uses the key and the rights
attributed to the key at its will. All keys are generated locally by their owners.
There is no hmit on the number of keys one physical entity may own. On the
other hand, if a physical entity is to receive any rights from others, it must be
represented by at leEist one key. Most public-key infrastructures (e.g. X.509) are
identity-oriented. In them, access rights are given to names of entities and the
names are separately bound to keys.
Delegation certificates also differ from traditional access control schemes in
that any key may issue certificates. There is no central or trusted authority that
could control the flow of access rights. All keys are free to delegate access to
services in their control.
The delegation takes effect only to the extent that the issuer of a certificate
itself has the authority it is trying to delegate. Nevertheless, it is perfectly legal
to issue delegation certificates for rights that one does not yet have or for broader
rights than one has in the hope that the issuer may later obtain these rights.
Sometimes a key may delegate all of its rights to another key. The policy for
calculating the access rights received by the the subject when the issuer itself
does not possess all of the rights listed in the certificate depends on the type of
authorizations in question. In this paper, we consider only set-type authorizations
as most literature [4,14]. The subject gets the intersection of the rights held by
the issuer and the rights mentioned in the certificate. Since this limitation is
 214

true for all delegation certificates, it is usually not explicitly mentioned. If we

also ignore the validity period, the certificate above can be written, in short,

SK{K' has the rights R.)

Since the signing of a certificate happens locally at the physical entity pos-
sessing the private issuer key, the act of signing does not invalidate any existing
certificates or affect existing rights. The subject of the certificate gets new rights
but the issuer does not lose any.
In this way, delegation is less powerful than transfer of rights where the
originating entity loses what it gives. On the other hand, delegation is far easier
to implement. The simplicity of the implementation (signing a certificate) in a
distributed environment is the reason why delegation is preferable to transfer as
an atomic access control primitive. (See Sec. 4.1 about implementing transfer.)

2.2 Certificate Chains

This section shows how the certificates are used as a proof of authority and how
the rights can be passed forward through several keys and certificates.
We begin by considering the simple case of access right verification where the
owner of a service has issued a certificate directly to the user of the service (like
K delegates to the public key K' above). When the user wants to use its rights,
it signs a request with its private key {K') or in some other way authenticates
(e.g. by estabhshing an authenticated session) the access request with the private
key. This can be construed as redelegating the rights to the request. The user
attaches the delegation certificate to the access request and sends both to the
server.
Every service platform has either a single master public key that controls all
access to it or access control lists (ACL) that determine the privileged public
keys for each service. When the server receives an access request with an at-
tached delegation certificate, it first verifies that the certificate is signed by a
key controlling the requested service (K). It then checks that the authorization
in the certificate is for the service and that the key requesting the service is
the same as the subject of the certificate {K'). In this scenario, the certificate
behaves like a capability. The signature protects the capability from falsification
and binds it to the subject key.
Just as a key may delegate rights to services it directly controls, it may also
redelegate rights it received by delegation from other keys. In this paper we
assume that redelegation is always allowed unless a certificate explicitly forbids
it.
When a key delegates to another key and this key in turn redelegates to a
third one, and so on, the delegation certificates form a chain. In the chain, the
access rights flow from issuers through certificates to subjects. The original issuer
is usually the service producer and the final subject is a client of the service.
If all the certificates in the chain delegate the same access rights and specify
the same validity period, these rights are passed all the way from the first issuer
 215

to the subject of the last certificate. But it is not necessary for the certificates
to have the same authorization field and validity period. We remember that the
rights obtained by the subject key are the intersection of the rights possessed
by the issuer and the authorization field of the certificate. Consequently, the
rights passed through the chain of certificates can be computed by taking the
intersection of the rights possessed by the first issuer with the authorizations
on all the certificates in the chain. (Sometimes the intersection can be empty
meaning that no rights are passed all the way.) Likewise, the validity period
of the of the chain is the intersection of the periods specified on the individual
certificates. For example, in the chain of Fig. 2, Key4 receives the right to the
web pages "http://S/file" for today only from key Keyl.

-Delegation certificate - -Delegation certificate — -Delegation certificate

Key2 may read web Keys may read web Key4 may read any
server S this year. pages http://S/file always. web servers today.
Signed: Key1 Signed: Key2 Signed: Key3

attach certificates to request

/v\ /^y]
-Access request-
Read http://S/file.
Server S Signed: Key4 Client
controlled by Key1 owner of Key4

Fig. 2. A chain of delegation certificates

Since the certificates can be issued by anyone to anyone, they do not neces-
sarily form simple chains. Instead, the certificates form a graph structure called
delegation network. In the delegation network, there may be many chains of cer-
tificates between the same pair of keys. Naturally, the rights passed between two
keys are the union of the rights passed by all individual chains between them.
(Set-type authorizations are combined with the union operator. Other policies
are possible for other types of authorizations.)
When a key requests a service and it has obtained the access rights through
a chain of certificates, it attaches the entire chain to its request (Fig. 2). The
server will verify that the chain originates from a key controlling access to the
requested service, that it ends to the key making the request, that each certificate
in the chain is signed by the subject of the previous certificate, and that each
certificate in the chain authorizes the request. Several chains of certificates can
be attached if the combined rights delegated by them are needed for the access.
The signing of an access request can also be thought of as redelegation to
the request. Furthermore, the request may be program code. In that case, it is
natural to think that the last key in a chain redelegates to the code and the code
makes the actual requests. Consequently, delegation certificates should be used
 216

to express this delegation. This is done so that the last key signs a delegation
certificate where the subject is not a key but a hash value of the program code.
We will see an example of this in Sec. 2.4.
In a way, the delegation certificates behave like signed requests for capability
propagation (see e.g [17]). The requests are honored only if the signer itself has
the capability. However, it is not necessary for the server or for a trusted party
to process each propagation before the next one is made, and no new capabilities
need to be produced before the rights are used. Instead, the information is stored
in the form of the delegation certificates.
Although the most obvious way of managing certificates is to accumulate
them along the chain of delegation and to attach them to the service requests,
there is no compelling reason to do this. The certificates can be stored and
managed anywhere as long as the verifier gets them in the end. The accumulation
is not always even possible if the certificates are not issued and renewed in the
order of their positions in a particular chain.
Managing long chains of certificates can become a burden. A technique called
certificate reduction saves work in verifying the certificate chains. We observe
that two certificates forming a chain

SxiiK'z has the rights Ri.) and 5/^2(i^s has the rights i?2-)
imply a direct delegation from Ki to K3:

SKI (KS has the rights Ri n i?2.)

The third certificate is redundant and signing it will have no effect of the access
rights of any entity. Hence, Ki can sign it after verifying the chain of signatures.
Continuing the process inductively, a chain of any length can be reduced into a
single certificate. If the reduced certificate is used repeatedly as a proof of access
rights, savings in the verification time can be substantial. In order to compute
the intersection of the authorizations, the issuer of the reduced certificate must
have some understanding of the contents of the authorization fields. Other than
that, the reduction is a mechanical procedure that does not involve any decision
making. Certificate reduction is the main technique of certificate management
in SPKI.

2.3 Distribution of Authority and Operations

This section explains the rationale behind delegation with certificates. We em-
phasize suitability for open distributed systems with no globally accepted au-
thority. The advantages center around the high level of distribution achieved
both in authority and management work load.
The first key to the distribution is that the entities are represented by their
cryptographic keys instead of names. Names are a natural way of specifying
entities for humans but they are less suitable for cryptographically secure au-
thentication. Ellison [13] discusses the complicated connection between keys and
identities. In a key-oriented world, we don't need trusted third parties to certify
 217

the binding between a name and a public key. Instead, the pubUc key is used
directly to specify an entity. Thus, the centralized or hierarchical certification
authorities (CA) that are the heart of traditional identity-based access control
lose their role in key-oriented systems. This is a major security advantage. For
example in X.509, the name certificates come from a global hierarchy of offi-
cials (CAs) who must all be trusted with respect to any access control decision
whose security depends on the correct mapping between a key and a name.
Key-oriented access control avoids such obvious single points of failure. Fig. 3
illustrates the difference. It becomes even more obvious when we remember that,
in a truly name-oriented system, the mapping from name to key must be done
also for the service owner and for every key in a chain of delegation.

Identity-oriented Key-oriented

Owner of service S,

Certification autfiority

Service S.

Service S.

Fig. 3. In an identity-oriented system, security of all access control depends on trusted

authorities. Key-oriented systems avoid this.

The omission of names also results in savings in communication and cer-

tificate processing. There is no need to contact a CA. The verification of the
certificates in a chain is a straight-forward, mechanical procedure that involves
no communication or policy decisions.
Sometimes we, nevertheless, need bindings between names and keys. This
is usually because of the need to refer to legal persons or because names are a
natural form of human input. In such cases, we must either trust name certificates
from a some CA or, preferably, resort to SDSI-type linked name spaces [1,24]
whose relativity is explicit. But when feasible, the most robust practice is to
pass the correct public key from the subject to the issuer at the time when a
delegation certificate is created and to issue the certificate directly to the key.
The effect of delegation on the distribution of authority, however, is greater
than only eliminating the CA. Delegation-based access control does not con-
 218

ceptually differentiate between keys that are allowed to grant access and ones
that only use services. From technical point of view, all keys are equal regard-
less of the importance of rights they handle. Any key may issue certificates to
others and distribute access rights to the services in its control without asking
permission from any other entity.
The bottom-up formation of policy is the most distinguishing property of
certificate-based access control. The certificates are issued locally by the entities
that produce the services and, therefore, should be responsible for granting ac-
cess to them. The certificates are formal documents of local trust relationships
that arise from voluntary personal, technical and business relations between the
entities possessing private keys. Nothing is mandated by global authorities. In a
chain of delegation certificates, every link is a result of a local policy decision.
Because of redelegation, the local decisions by individual key owners have global
consequences.
The lack of enforced hierarchical structure makes the system open and scal-
able. Setting up new a new entity is as simple as generating a signature key. New
keys may be created locally as new physical entities or services are introduced.
In comparison, most traditional access control systems achieve scalability with
a hierarchy of trusted entities or domains and require meta-level maintenance
operations for changes in the hierarchy [12,16,18].
An important observation in the certificate management is that the storage
and distribution of certificates is a separate concern from their meaning [8,9].
The integrity of the certificate data is protected by the signatures. Thus, un-
trusted entities can be allowed to handle them. Often, organizations will want
to set up certificate databases for access-right acquisition and management. Sim-
ilarly, most application software packages are unlikely to include their own cer-
tificate management. Most of the tasks can be done for them by untrusted servers
or helper software. Nikander and Viljanen [23] describe a way of storing certifi-
cates in the Internet domain name service and a discovery algorithm based on
[3].
The certificates should be considered as sets or graphs rather than as chains.
This is because the order of issuance of the certificates does not necessarily have
any correlation to their order in a particular chain. The time of issuance and the
vahdity periods depend on the local trust relations behind the certificates. These
are independent of whatever chains may be formed globally. If one certificate in
a chain expires, only that one needs to be refreshed immediately. The other
certificates in the chain remain valid. In this way, the system is distributed not
only in space but also in time. It is even possible that the private keys or the
owners of the keys who issued some of the certificates have ceased to exist by
the time the certificates are used as a part of a chain. This is commonly so when
temporary key pairs are used for anonymity or when a old system delegates its
tasks and rights to a replacing one.
Certificate reduction allows a trade-off between communication and certifi-
cate processing cost. Reduction requires one to contact the issuer of the reduced
certificate. This means that the entity must be on-line at the time of the re-
 219

duction and that this on-hne system must be secure enough to hold the private
signature key. Luckily, the reduction engine can be fairly simple to build. If all
the certificates in the chain delegate the same rights or if the application has
straightforward rules for computing intersections of the authorizations, the re-
duction is a purely syntactical transformation and it can be fully automated. The
issuer of the reduced certificate does not need to fully understand the meaning
of the certificates. The reduction may save significantly in the costs of certificate
transfer and verification.
Unlike mandatory access control mechanisms, delegation does not have a
central reference monitor to supervise access and distribution of the access rights.
Neither is there a trusted computing base (TCB) to monitor the actions of the
distributed entities. This is because there is no global policy to enforce on the
parts of the system. In open environments like the Internet, it would not be
possible to implement any global controls. Instead, the policy is determined
locally by the parts.
Delegation leaves more to the discretion of the participants than many other
DAC mechanisms. The system does not enforce any policies to protect users from
bad decisions. All the verifier of a certificate sees is public keys and signatures.
It has no way of telling if the private keys belong to legitimate entities in the
system. The certificate issuer at the time of signing should, of course, have a
soHd reason for trusting the subject with the delegated rights. But the systems
leaves it to each key owner to judge by itself who can be trusted. The issuer
may or may not know the name or identity of the subject key owner. Moreover,
redelegation creates a new degree of freedom. When redelegation is allowed,
anyone can share his rights with others. This is equal to universal grant rights
from anyone to anyone.
It may seem that redelegation should be controlled. It is, indeed, possible to
add conditions on the certificates hmiting redelegation. In SPKI, the choices are
to allow or forbid redelegation completely. A certificate that forbids redelegation
can only be the last certificate in a certificate chain.
There are, however, appealing arguments for allowing free redelegation. It
may be, convenient for the client to authorize someone else to use the rights on
its behalf, or the client may want to redelegate the rights to one of its subsidiaries.
The internal organization of the clients of a service should not be a concern to
the service providers. Free redelegation makes the internals of the system parts
more independent thus furthering distribution.
The key-oriented nature of the system also obscures the semantical meaning
of the restrictions on redelegation. Only in special circumstances does the issuer
of a certificate know that the private part of the subject key is held permanently
secret by a certain physical entity. In many cases, the issuer accepts any key given
by the entity that is to receive the rights. In general, there is no guarantee that
the corresponding private key will not be revealed to others. Giving out the key
would spread the rights as effectively as redelegation but, unlike in redelegation,
the rights and their validity period could not be limited. Usually, there are also
 220

other ways to redistribute the services without the agreement of the originator,
e.g. estabhshing proxy servers and outright duplication of the server data.
Instead of forbidding redelegation, the issuer of a certificate should consider
changing the authorization and validity period. Minimizing the scope of dele-
gated rights is the most natural way to express limited trust in the subject. The
certificates make it easy to reconsider the rights and validity in each step of a
delegation chain.
All in all, delegation does well the part of access control that is easy to im-
plement: maximally discretionary distribution of access rights. Mechanisms for
identity certification, limits on redistribution, rights transfer and revocation can
be added where they are required. However, such features in principle require
a more complex infrastructure with a TCB, tamper-resistant modules, trusted
third parties or on-line communication. Therefore, the basic access control sys-
tem should not require their use. We believe that there are many instances where
pure delegation is a sufficient mechanism and corresponds well to the real-life
access-control needs.

2.4 Access Control for IN Services

Delegation certificates are most suitable for use in distributed systems with no
globally accepted security policy or authority and in open systems with no central
registration of the servers and clients. They are unlikely to find apphcations in
high-security environments where mandatory access control and trusted systems
are a rule. Although most apphcations will be on the Internet, we will look at
an example from the telecommunications world.
Calypso [19] is a distributed service architecture for intelligent networks (IN).
It is designed for ATM access networks where a workstation running the Calypso
service platform controls an ATM switch. Calypso provides flexible distribution
of service and network control functions among service clients, servers and net-
work nodes. The same architecture could also reside on top of other types of
network equipment such as IP switches and firewalls.
Calypso is based on a business model where the network operator who owns
the infrastructure offers network resources to service providers (SP). These re-
sources are the lowest level of Calypso services. Service providers can either
market the right to use their services to end users or they can sell them to other
SPs for reselling or for use as building blocks of more sophisticated services.
Complex services and their components form tree-like structures (Fig. 4).
All Calypso services are implemented as Java pacltages. A service may use
other services by calling methods of the classes that belong to them. Hence, the
network nodes must have an access control system that facilitates execution of
code from mutually distrusting SPs and contracting of services between them.
The Calypso security requirements and a tentative architecture for satisfying
them were outlined in [6].
The IN access control mechanisms should encourage free formation of busi-
ness relations between service producers and merchants distributing access rights
 221

(ServicelJ SP1

' Delegate
\ Service2 .'
A Delegate
I Service4
SP2 Delegate
Services

[services] (Service4j A
Delegate
SP3
\
SP4
\l
[services)
Services
SP5 A
Delegate
V
Delegate
Services
I
/
Delegate
Services I Services
\l V \L \ I /
(a) [ Low-level networ[< services J (b) Network operator

Fig. 4. Calypso service composition and delegation between SPs

to the services. Delegation, in a natural way, allows complex relations between

the entities without forcing too much predefined structure on the service market.
Calypso uses SPKI certificates to delegate access rights between SP keys. The
Calypso system differs from the standard scenario in that the last certificate
in the chain always delegates the rights to (a hash value of) a Java package
implementing the service that will use the access rights.
Fig. 5 shows how a service provider delegates access rights for its service
to another SP that in turn delegates them to its service code. (This is the
circled delegation step in Fig. 4(b).) When the code package is installed into
the network nodes, it is accompanied by the certificates. In Fig. 6, the access
rights are delegated through a service broker adding another link to the chain
of certificates.
The delegated authorizations are always rights to access a Calypso service.
Since services in the Calypso architecture are encapsulated into Java classes and
packages, this gives a natural level of granularity for the access control.
Each service has exactly one key associated with it: the key of the service
author or owner. This key distributes the access rights by issuing delegation
certificates. The decision to delegate depends on the relations between the enti-
ties possessing the private keys. However, the service producers do not actually
need to think in terms of access control policies. Instead, they sell and buy the
certificates with business relations in mind.
The network operator or a service provider may want to review the client
service implementation before granting access. This is because the access control
mechanisms cannot effectively prevent denial of service, bad publicity and many
other potential problems created by malicious or low quality services. The review
 222

— Delegation certificate -
Service Service
provider 2" Keyl may access Service2. ^ provider 1
holds private Key2. Signed: Key2
holds private Key1.
I
~ Delegation certificate'
hash(Service1 code)
author or
may access Service2.
owner
Signed: Keyl

: ^ Service platform
Service2
owned by Key2. Servicel

F i g . 5. Delegation to another service provider (SP) in Calypso

Service
broker
holds private KeyB.
\
/ — Delegation certificate -
— Delegation certificate - Key1 may access Service2.
KeyB may access Service2. Signed: KeyB
Signed: Key2

Service / \ Service
provider 2 provider 1
holds private Key2. holds private Keyl.

F i g . 6. Delegation through a service broker in Calypso

 223

process is not easy to organize because the aim is to allow fast development of
services by a large number of independent SPs. A possible solution is to use
independent quality-control (QC) units that certify services if they meet some
minimum quality criteria. In Fig. 7, the network operator grants access rights to
SPl's code only after receiving the review results. If the QC writes a certificate
to Keyl instead, that means it has reviewed the production process of SPl and
the network operator can trust SPl to do its own quality control.

Network
operator
holds private KeyN.
1 — Delegation certificate —
Network operator must see hash(Service1 code) may
the quality certificate from access Network services.
Signed: KeyN
KeyQ before signing.

Deiegatlon certificate -
Quality tiash(Service1 code)
control passed quality check.
liolds private KeyQ. Signed: KeyQ

Fig. 7. Code quality control for IN services with basic delegation certificates

The example reveals a major limitation of the basic delegation certificates.

The quality check must be passed before the network operator makes its decision.
If changes are made to the code, the network operator must renew its certificate.
But the network operator clearly does not make any new policy decision at that
time. It only follows a fixed rule: if the quality check is ok, it will sign. In the
following sections we will see how the requirement for quality check and other
rules for deriving new access rights from old ones can be encoded into certificates.

3 Threshold and Conditional Delegation

The following sections introduce certificates with more complex structure than
the ones we have seen so far. The enhancements increase significantly the fiexi-
bility of certificates as an access control tool. Threshold certificates are a means
of dividing authority. Instead of giving the rights to a single subject, they are
given to a. group of subjects who must co-operate to use the rights. Sec. 3.1
describes the certificates and Sec. 3.2 explains how they are used. Conditional
delegation is a way of expressing simple access control policy rules in certificates.
 224

We will introduce the new type of certificates in Sec. 3.3 and look at applications
in Sec. 3.4.

3.1 Threshold Certificates

Threshold certificates are an extension of the basic certificate structure. Instead

of having only a single subject, they divide the authorization between several
subjects. In order to use the rights given by a certificate, the subjects must co-
operate. The certificate has a threshold number that determines how many of
the subjects must agree.
A (A;,n)-threshold certificate is a signed message like the following:

S'ii'(During the validity period Ti — TJj, if I have any of the rights R,

I give them also to {Ki,K2-, •••, Kn},
k of whom must co-operate to use or redelegate the rights.)

The usual way for the subjects to co-operate is to redelegate the rights to
one of them or to some other single entity. In Fig. 8, KeyC receives the right
R from KeyS because two subjects of the (2,3)-threshold certificate co-operate
to pass it to KeyC. When KeyC wants to use the right R, it must attach all
three certificates to its access request. It should be noted that the subjects of
the threshold certificate do not need to delegate directly to the same key as in
the figure. The delegation could go through independent or partially dependent
chains of certificates and even through other threshold certificates before the
shares are accumulated to a single key. The general structure of such networks
was studied in [4].
The threshold certificates can also be used in some situations where there
is no real threshold trust scheme. An example below will show how they may
improve distribution and flexibility of the system.
(A;,A;)-threshold certificates where all subjects are required to co-operate are
sometimes called joint-delegation certificates. Open threshold certificates [4] are
a variation of the threshold certificates where each subject is given a separate
certificate and new subjects can be added later.

3.2 Threshold Certificates and IN Access Control

There may be intelhgent network services controlled by a group of SPs so that

two or more are needed to grant access to the services. Such conditions can be
directly expressed with threshold certificates from the service master key to the
service providers' keys. A single SP may also want to ease the burden of storing
its private key securely by distributing the authority between several keys. This
is exactly what is pictured in Fig. 8. The SP generates three new keys Ki, K2
and K3. It then signs a the following kind of certificate with its master key K.

5K(Any 2 of {Ki,K2,K3} have all my rights.)

 225

Service master key,

> ^ erased after signing.

— (2,3)-threshold certificate
Any 2 of {Keyl ,K6y2,Key3}
fiave all my rights.
Signed; KeyS

y
Key1
?
Key2
^ Keys
Keys that share the
autfiority of KeyS.
I
Delegation certificate
y
KeyC has right R. Delegation certificate
Signed: Key2 KeyC has right R.

T/
KeyC
Signed: Key3

Fig. 8. A threshold certificate {KeyC gets right R from KeyS)

This certificate allows any two of the new keys to operate on behalf of the
original master key. The new private keys are stored in separate places while the
old private master key is destroyed or stored in a safe place and never accessed.
This protects both against theft and accidental loss of the private master key.
If one share is compromised, it alone is not enough to misuse the service, and if
one share of the three is lost, the other two can still grant access to the service.
The SP can still advertise its old public key {K or KeyS) and receive new rights
delegated to that key. The threshold certificate passes all these rights to the
share keys who can authorize others. This kind of protection of private keys
may prove to be a much more common reason to use threshold certificates than
actual threshold trust schemes between business associates.
There is another, rather unexpected, application for the threshold certificates.
We will see that flexibility can be added to systems like that of Fig. 7 by encoding
an imphcation rule in a certificate. If Fig. 9, the quality-control key {KeyQ)
certifies the code by granting the code all its rights. The network operator issues
a (2,2)-threshold certificate for SPl and QC. Thus, SPl needs the agreement of
the QC before it can use the network services. The three certificates together
convey the right to access Network services from KeyN to Servicel. With these
certificates, Servicel can prove its access rights when installed into a network
node.
The QC key is never used for any other purpose than for certifying entities
that have passed the quality check. A quality certificate can be issued to code
or to service providers whose own quality control has passed an audit. In the
 226

quality certificates, the QC delegates all its rights so that it does not need to
know what kind access the quality certificates are used for. This means that the
QC key should never be given any other rights than one share of a threshold
certificate for the purposes of a quality check on whoever gets the other share.

Network Service
operator provider 1
holds private KeyN. holds private Key1.

\ ^
- (2,2)-threshold certificate - — Delegation certificate —
Key1 and KeyQ together may hash(Service1 code) may
access Network services. access Networl< services.
Signed: KeyN Signed: Keyl

Quality
control
~y. Delegation certificate -
hash(Service1 code)
?
has all my rights. Servicel
holds private KeyQ. Signed: KeyQ

Fig. 9. Code quality control for IN services with threshold certificates

Compared to the use of basic delegation certificates (Fig. 7), the threshold
certificate has the advantage that dependences between the quality control and
the granting of access to the services have been reduced to minimum. It is not
necessary to involve the network operator every time code is changed and the
certificates can be signed in any order. The three certificates can be issued and
renewed independently of each other.
Admittedly, this is a somewhat inelegant way to use threshold certificates.
The rights delegated by the certificate are not encoded in the the authorization
part ("all my rights") but in the signing key {''KeyQ is the quality-control key").
If there are several different authorizations (e.g. several types of quality check),
the QC unit must have an equal number of signature keys. Moreover, comparing
different levels of authorization becomes difficult when the authorizations are
encoded in the keys. We observe that the threshold certificate in Fig. 9, in effect,
carries the meaning
"Keyl may access the network services if it also has
a quality certificate from KeyQ."
In section 3.4, we will introduce a new type of certificate that explicitly in-
cludes such conditions. The reason why we have described in length how to
encode the same meaning into a threshold certificate is that only threshold del-
egation is currently supported by SPKI.
 227

3.3 Conditional Certificates

Conditional certificates are like the basic delegation certificates except that they
state additional conditions that have to be satisfied before the certificate is con-
sidered valid. A conditional certificate is a signed message with the following
contents:

5iv: (During the validity period P, if I have any of the rights R,

I give them also to K' if it also has
the right Ri from Ki, and
the right R2 from K2, and
the right R3 from K3, ...)

The certificate gives the rights to the subject only if all the conditions in the
list are satisfied. The conditions always take the same form: they require the
subject key of the certificate to have a certain authorization from a certain key.
This is natural because any attribute that the subject may have can be verified
only if it is expressed as an attribute certificate from a proper authority.
In order to use the certificate, the subject must provide a proof that the
conditions are fulfilled. It does this by attaching appropriate certificate chains,
one for each condition. The certificates in the proof of access rights form a tree
(or a directed acyclic graph) rather than a chain.
When certain attributes are required before granting access to a service, a
conditional certificate offers two advantages compared to the basic delegation
certificates. First, the certificate is an unambiguous, standard-form statement of
what kind of attribute certificates are still needed for the access. Secondly, the
conditional certificate can be signed before obtaining the attribute certificates.
Without conditional certificates, the client in need of access rights would first
contact the issuer to find out what are the prerequisites, then try to acquire them
and, in the end, return to the issuer with the collection of credentials in order
to get the new certificate. When the decision rule is encoded in a certificate,
the issuer needs to be contacted only once. The client may obtain the attributes
before or after this contact. Thus, communication and synchronization between
the entities is greatly reduced.
The conditional certificates express simple policy rules but they are, by no
means, a general language for defining policies. For example, the certificates pre-
sented here cannot express symbolic rules. A more general language for express-
ing conditions, policies and limits on redelegation is a topic of active research.

3.4 Conditional Certificates and IN Access Control

Conditional certificates are just the right tool for the kind of situations where we
shghtly abused threshold certificates in Sec. 3.2. The code quahty check for an
IN service can be expressed as a condition. Fig. 10 reformulates the certificates
of Fig. 9 with conditions. The result is functionally the same but the system is
much more intuitive for a designer or an observer.
 228

With the threshold certificates, the quahty certificate told only indirectly
what kind of authority it gives to the subject. The authorization was encoded
in the signature key. In addition to being cumbersome to understand, this en-
coding has the disadvantage that the authorizations are not comparable. The
conditional certificates, on the other hand, explicitly state the rights or attributes
in the authorization field of the certificate and the authorizations can be com-
pared. For example, we can let the IN Network operator write the required level
of quality in the conditional certificate. If the QC issues a quality certificate with
the same or higher level to Servicel, the code will get the access rights.

- Conditional certificate — i
Network _ Service
operator
Keyl may access Network ' provider 1
service if it has "passed holds private Keyl.
holds private KeyN quality check" from KeyQ.
Signed: KeyN

- Deiegation certificate —
hash(Service1 code) may
access Network service.
Signed: Key1

Quality
control
Delegation certificate -
hash(Service1 code)
T
passed quality check. Servicel
holds private KeyQ. Signed: KeyQ

Fig. 10. Code quality control made simple with conditional certificates

4 Limitations
The flexibility of the certificates does not come completely without a price. There
are many security goals that require centrahzed control and cannot be realized
only with signed messages. We will consider such goals and see what kind of
central or trusted services they imply. Sec. 4.1 discusses policies that require
additional infrastructure. Sec. 4.2 brings up the issue of quantitative rights. In
Sec. 4.3 we consider revocation and in Sec. 4.4 anonymity and auditing.

4.1 Expressive Power

The certificates are a form of discretionary access control. They cannot express
mandatory policies like the Bell-LaPadula model [7] because we do not assume
 229

any mechanism for enforcing a policy globally. A mandatory policy would require
all equipment on the system to be under the control of some authority so that
they can be trusted to follow the pohcy.
Another limitation is that the certificates can only convey policies where
the rights of the entities grow monotonically as they acquire new certificates. It
is impossible to verify that someone does not have a certificate. Consequently,
separation-of-duty policies like the Chinese Wall policy [11] cannot be expressed
with only certificates. They need some mechanism for keeping track of the previ-
ously granted rights. Moreover, if several distributed issuers give out certificates
for different conflicting rights, these issuers must share a single view of the sub-
jects' histories. The histories must be updated in real time when new certificates
are issued. An equally difficult problem is that in a key-oriented system, one
physical entity might use several keys to gain conflicting rights. We must first
identify the entities whose duties we want to separate and then find a way of
mapping keys to unique identities. For example, a trusted official could certify
the keys to be unique personal keys of the participating persons. Altogether, sep-
aration of duty appears to be one of the greatest challenges for certificate-based
access control.
A related problem is the separation of access rights and grant rights. In a
key-oriented architecture, someone with only grant rights could easily subvert
the protection mechanism by issuing the rights to a key held by himself [15].
Therefore, every key in practice has the rights that it is allowed to delegate to
others. Like the separation of duty, separate policies for granting and using rights
cannot be securely implemented unless each entity has a unique identity and
keys owned by the entity are bound to the identity. Consequently, key-oriented
systems usually do not even try to implement pure grant rights.
An occasionally needed access control feature is a proxy that can issue cer-
tificates with longer life-times than its own existence. For example, a manager
on a vacation should be able to delegate authority to a stand-in only for the time
of the leave but the decisions made by the stand-in should stay valid longer than
that. Unfortunately, the proxy can create valid-looking certificates even after
losing its authority. He simply writes false dates on them so that they appear
to be signed at the time when he was authorized. The problem can be solved
with the help of a trusted time-stamping service if such a centralized authority
exists. Often it is easiest to let the certificates expire when the mandate of the
proxy does and have the master entity revalidate them.
The above limitations are due to fundamental properties of the access control
mechanism. There are, however, some other respects in which we have deliber-
ately satisfied us with less than the maximal expressive power. For example,
symbolic expressions could be allowed in conditional certificates. The extension
would make it possible to express general rules while the certificates proposed in
this paper can only speak of fixed keys. The reason for presenting the less gen-
eral model here is that it solves the practical problems with the basic delegation
certificates that we have met in applications. Future work on symbolic condi-
tions will determine the extent to which they are worth the increased complexity.
 230

The question of the optimal expressive power for the certificates involves issues
of computational and communication complexity and typical usage patterns in
applications.
A question that we will leave open is the exact structure of the authoriza-
tions. The types of access rights and the policies for combining them depend
on the application. In Policy Maker [8,9], the authorizations are expressed as
small programs of a safe programming language and certificates can communi-
cate with each other. This maximally generic approach leads to concerns about
the tractabiUty of access control decisions [10]. But no matter how the authoriza-
tions are encoded, certificates have one general limitation in this respect: they
can effectively express qualitative authorizations but not quantitative. That is
the topic of the next section.

4.2 Accounting and Redelegation

Accounting of service usage is an essential function in many commercial appli-
cations. The delegation certificates themselves do not support accounting in any
way. They are inherently reusable and can be combined with any number of
access requests without losing their validity. For example, there is no use-once
certificate or quota on the number of times a service can be used. Such fea-
tures can, of course, be built with additional infrastructure but they cannot be
encoded into certificates.
If accounting is difficult, a natural alternative is to charge a flat rate per
client. But as we argued in Sec. 2.3, it is an equal challenge to keep the subjects
from sharing their rights with others. By sharing their access rights, the clients
can frustrate the flat rate charging policy.
The reason behind these problems is again the lack of unique identities for
the individuals in the system and the lack of a global authority that could police
the actions of the individuals. There are three main approaches for overcoming
the problems with accounting and charging:

1. Make the charging recursive. Every reseller will be responsible for collecting
payments from the clients it delegated rights to or for dividing quotas be-
tween them. Apart from physical control of the clients, there are two ways in
which the reseller can divide the services and the costs between its clients.
(a) The reseller divides the service capacity at its disposal into smaller time
slots and more specific methods of access. Because the authorizations
are refined to be suitable only for very narrow purposes, the clients must
repeatedly request new certificates from the reseller who collects usage
data.
(b) The authorized clients may use the services at any time and the server
collects usage data. The usage statistics and the certificate chains that
were used as proof of access rights are propagated from the server down
the tree of resellers.
2. Require the client keys in the system to be certified by a trusted entity that
guarantees their payments or gives them a credit rating. The server verifies
 231

the credit before allowing access and collects payments directly from the
clients. This may not stop the sharing of access rights but it means there is
someone to pay for the metered usage.
3. Require the participants to incorporate a tamper-resistant police module on
their systems. Only keys on the tamper-resistant modules are allowed to
participate in the distribution of the access rights. The module can do the
accounting or enforce whatever limitations are wanted.

All these techniques incur a cost in that they require additional infrastructure
and make parts of the system less independent. But these costs are inherent to
any accounting and charging mechanism. With delegation certificates, we can
decide separately for each application if such measures are needed and if their
cost is acceptable.
In Calypso, we have chosen the approach 1(b) because the delegation chains
between SPs are relatively short (often only one step) and the charging for
component services is arranged in the same tree-hke manner. Access rights and
payments flow in the opposite directions in the service composition tree (Fig. 4).
A problem related to accounting is access rights transfer where an entity giv-
ing rights to others loses them itself. From a chain of certificates, it is not possible
to see if the chain ends there or if the rights have been redelegated further. Thus,
an entity that has redelegated its rights can still use them. Implementing trans-
fer in a distributed system requires a TCB or tamper-resistant police modules.
(The tamper-resistant module can, in fact, be thought of as a TCB.) One such
system for the transfer of software licenses between tamper-resistant smart cards
is described in [5].

4.3 Certificate Revocation

Sometimes an entity distributing access rights may want to reverse its decision
after the rights have already been granted. The change in mind may be due
to changed circumstances or more accurate information about the subject. In
a certificate-based architecture, this means invalidating certificates after they
have been issued but before their expiration dates. In general, any decrease in
the trust placed on the subject may require the issuer to sign and new certificate
and to cancel the old one.
In systems where all access is controlled by one reference monitor, access-right
revocation is a simple matter. It suffices to update the access control lists or to
store information on the exceptions at the place where the rights are verified.
In a distributed system, ACLs are stored and decisions to grant access are
made in more than one place. It is necessary to propagate the information on re-
voked access rights to all these places. The communication causes unpredictable
delays and, consequently, real-time revocation cannot be achieved like in a cen-
tralized system. An efficient infrastructure for propagating the revocation infor-
mation is a central part of many access control systems (for example [17,21]).
Some options are to broadcast revocation events or to notify only the interested
 232

parties, to immediately propagate every single revocation or to periodically up-

date exception lists, and to transfer the information in push or pull fashion.
All the revocation methods add complexity to the access control architecture.
They imply frequent communication and on-line presence of servers and clients.
On open networks, it would be unreasonable to require all entities to set up the
infrastructure for efficient revocation. Therefore, we prefer to avoid revocation as
far as possible. Instead, we suggest limiting the validity time and authorization
of a delegation certificate. They should reflect the level at which the subject can
be trusted not to misuse the rights. If revocation is needed frequently, it may
be just as feasible to require frequent refreshing of the certificates. The TAOS
[25] operating system is an example of an architecture that relies completely on
expiration dates instead of revocation.
If revocation is nevertheless needed, the implementation should follow two
principles. First, the cost of the additional infrastructure should be paid only by
the entities that want the service and only when it is strictly needed. If the issuer
of a certificate wants to be able to revoke it, he should be responsible for setting
up the distribution channels for the notification. This may be a simple task if
the certificates for the rights in question are always verified in a fixed set of
servers and extremely complex if there are unregistered off-line verifiers. Second,
no system should completely abandon expiration times in favor of revocation. If
expired certificates are not purged from the revocation lists, the lists can grow
uncontrollably. The more frequently revocation occurs, the shorter the validity
periods should be. Finding the right balance that minimizes the costs requires
careful analysis of each individual system.
SPKI certificates can include instructions for downloading a lists of revoked
certificates or for asking confirmation from an on-line server. If the issuer of
a certificate wants to be able to revoke it, he must explicitly state this in the
certificate. This way the cost of the revocation lists is paid only when it is strictly
needed. Moreover, there is no centralized system for maintaining and distributing
the lists. Instead, each certificate can have a reference to the particular server
and the signature key that provide the lists.
With conditional certificates, the confirmation from an on-line server can
be implemented in the same way as the quality check in Fig. 10. Normally the
validity period of the quahty certificates from the QC is long or infinite. Making
it very short would be the same as asking the SP to get a frequent on-line
confirmation from the QC.

4.4 Auditing and Anonymity

Designers of secure access control have traditionally emphasized audit capabili-

ties. That is, every action should be traceable to an entity that can be held re-
sponsible for it. The shift from closed mihtary networks to open commercial ones
has brought a conflicting need, privacy. Although auditability and anonymity
are both desirable security goals, it is not easy to achieve both simultaneously.
Hence, we must try to balance the requirements according to the application.
 233

The signature key does not directly reveal who is responsible for a signed
request which makes it difficult to trace the responsible parties. For auditing,
the keys must be bound to the persons or legal entities that are liable for their
actions. That kind of bindings can be created by identity-escrow agents that
guarantee to find a responsible person if the need should arise. The escrow agents
issue certificates to the keys whose identities they have escrowed. The services
that require auditing only accept clients with the escrow certificates.
The key-oriented system protects the users' privacy by not explicitly revealing
their names. However, the keys are easily recognizable identifiers that can be used
to combine data collected from different sources. Therefore, further measures
are needed for reliable privacy protection. The certificate reduction (see Sec.
2.2) helps in some cases. A chain of certificates may reveal the identities of the
intermediate entities but when the chain is reduced, that information is hidden.
SPKI puts great emphasis on privacy aspects and relies mostly on the reduction.
An alternative anonymity technique is to create temporary keys that do
not reveal their owner. When a subject entity wants its anonymity protected,
it provides the issuer of a new certificate with a freshly generated public key.
The temporary keys cannot be recognized and connected to the owner or to
each other. This is often preferable to certificate reduction because the entity
responsible for generating the temporary keys is the one whose anonymity is at
risk. Although the generation of the temporary keys is costly, it can be done
off-line in advance. With both techniques, however, the cost of privacy is an
increase in communication and synchronization between entities.

5 Conclusion
This paper described delegation certificates and some of their applications in
distributed access control. The goal was an abstract understanding of the basic
ideas without implementation details. We found that the main advantages of
the certificates he in decentralization. We also introduced conditional certificates
that help in further distribution of management operations in the system.
Delegation catches well the spirit of what is natural to access control of
distributed digital services. Some access control policies require additional in-
frastructure such as a TCB or trusted servers. We feel that such costs should
be avoided wherever possible. When their limitations are kept in mind, the del-
egation certificates can satisfy many every-day access control needs and can be
used as a uniform basis for distributed discretionary access control.

Acknowledgements
The work was funded by Helsinki Graduate School for Computer Science and
Engineering (HeCSE) and Academy of Finland. In am thankful to professor OUi
Martikainen and to Petteri Koponen and Juhana Rasanen for allowing the use
of Calypso as a case study. Part of the work was done while the author was at
UC Davis Computer Security Laboratory.
 234

References
1. Martin Abadi. On SDSI's linked local name spaces. In Proc. 10th IEEE Computer
Security Foundations Workshop, pages 98-108, Rockport, MA, June 1997. IEEE
Computer Society Press.
2. Martin Abadi, Michael Burrows, Butler Lampson, and Gordon Plotkin. A calculus
for access control in distributed systems. ACM Transactions on Programming
Languages and Systems, 15(4):706-734, September 1993.
3. Tuomas Aura. Fast access control decisions from delegation certificate databases.
In Proc. 3rd Australasian Conference on Information Security and Privacy ACISP
'98, volume 1438 of LNCS, pages 284-295, Brisbane, Australia, July 1998. Springer
Verlag.
4. Tuomas Aura. On the structure of delegation networks. In Proc. 11th IEEE
Computer Security Foundations Workshop, pages 14-26, Rockport, MA, June 1998.
IEEE Computer Society Press.
5. Tuomas Aura and Dieter GoUmann. Software license management with smart
cards. In Proc. USENIX Workshop on Smartcard Technology, Chicago, May 1999.
USENIX Association.
6. Tuomas Aura, Petteri Koponen, and Juhana Rasanen. Delegation-based access
control for intelligent network services. In Proc. ECOOP Workshop on Distributed
Object Security, Brussels, Belgium, July 1998.
7. D. Elhott Bell and Leonard. J. LaPadula. Secure computer systems: Unified ex-
position and Multics interpretation. Technical Report ESD-TR-75-306, The Mitre
Corporation, Bedford MA, USA, March 1976.
8. Matt Blaze, Joan Feigenbaum, John loannidis, and Angelos D. Keromytis. The role
of trust management in distributed systems security. In J. Vitek and C. Jensen,
editors. Secure Internet Programming: Security Issues for Distributed and Mobile
Objects, LNCS. Springer-Verlag Inc, New York, NY, USA, 1999.
9. Matt Blaze, Joan Feigenbaum, and Jack Lacy. Decentralized trust management.
In Proc. 1996 IEEE Symposium on Security and Privacy, pages 164-173, Oakland,
CA, May 1996. IEEE Computer Society Press.
10. Matt Blaze, Joan Feigenbaum, and Martin Strauss. Compliance checking in the
PolicyMaker trust management system. In Proc. Financial Cryptography 98, vol-
ume 1465 of LNCS, pages 254-271, Anguilla, February 1998. Springer.
11. David F. Brewer and Michael J. Nash. The Chinese wall security policy. In Proc.
IEEE Symposium on Research in Security and Privacy, pages 206-214, Oakland,
CA, May 1989. IEEE Computer Society Press.
12. Recommendation X.509, The Directory - Authentication Framework, volume VIII
of CCITT Blue Book, pages 48-81. CCITT, 1988.
13. CaxI M. Ellison. Establishing identity without certification authorities. In Proc.
6th USENIX Security Symposium, pages 67-76, San Jose, CA, July 1996. USENIX
Association.
14. Carl M. Ellison, Bill Franz, Butler Lampson, Ron Rivest, Brian M. Thomas, and
Tatu Ylonen. SPKI certificate theory, Simple public key certificate, SPKI examples.
Internet draft, IETF SPKI Working Group, November 1997.
15. Carl M. Ellison, Bill Franz, Butler Lampson, Ron Rivest, Brian M. Thomas, and
Tatu Ylonen. SPKI certificate theory. Internet draft, IETF SPKI Working Group,
October 1998.
16. M. Gasser, A. Goldstein, C. Kaufman, and B. Lampson. The digital distributed
system security architecture. In Proc. National computer security conference, pages
305-319, Baltimore, MD, USA, October 1989.
 235

17. Li Gong. A secure identity-based capability system. In Proc. 1989 IEEE Sympo-
sium on Research in Security and Privacy, pages 56-63, Oakland, CA, May 1989.
IEEE, IEEE Computer Society Press.
18. J. Kohl and C. Neuman. The Kerberos network authentication service (V5). RFC
1510, IETF Network Working Group, September 1993.
19. Petteri Koponen, Juhana Rasanen, and OUi Martikainen. Calypso service architec-
ture for broadband networks. In Proc. IFIP TC6 WG6.7 International Conference
on Intelligent Networks and Intelligence in Networks. Chapman & Hall, September
1997.
20. Ilari Lehti and Pekka Nikander. Certifying trust. In Proc. First International
Workshop on Practice and Theory in Public Key Cryptography PKC'98, volume
1431 of LNCS, Yokohama, Japan, February 1998. Springer.
21. Nataraj Nagaiatnam and Doug Lea. Secure delegation for distributed object envi-
ronments. In Proc. 4th USENIX Conference on Object-Oriented Technologies and
Systems (COOTS), pages 101-115, Santa Fe, NM, April 1998. USENIX Associa-
tion.
22. A guide to understanding discretionary access control in trusted systems. Technical
Report NCSC-TG-003 version-1, National Computer Security Center, September
1987.
23. Pekka Nikander and Lea Viljanen. Storing and retrieving Internet certificates.
In Proc. 3rd Nordic Workshop on Secure IT Systems NORDSEC'98, Trondheim,
Norway, November 1998.
24. Ronald L. Rivest and Butler Lampson. SDSI — A simple distributed security
infrastucture. Technical report, April 1996.
25. Edward P. Wobber, Maiti'n Abadi, Michael Burrows, and Butler Lampson. Authen-
tication in the Taos operating system. ACM Transactions on Computer Systems,
12(l):3-32, February 1994.
26. Philip Zimmermann. The Official PGP User's Guide. MIT Press, June 1995.
 A View—Based Access Control Model for
CORBA

Gerald Brose

Abstract. Specifying and managing access control policies for large dis-
tributed systems is a non-triviai task. Commonly, access control policies
are specified in natural language and later reformulated in terms of a par-
ticular access control model. This paper presents and discusses concepts
for an object-oriented access model that is more suitable for describing
access control policies for CORBA objects than the default access model
specified in the OMG security service specification.

1 Introduction

Specifying and managing access control policies for large distributed systems is
a non-trivial task. Commonly, access control policies are initially specified in
natural language and later reformulated in terms of a particular access control
model, e.g. a label-based or access matrix-based model. Because controlling the
dynamic evolvement of policies and adjusting them by administrative operations,
i.e. managing them, is highly sensitive and at the same time error-prone if done
at too low a level of abstraction, language support for this kind of activity would
be desirable.
As has been pointed out above, the language for specifying access control
policies is defined by the underlying access model. If the concepts and abstrac-
tions of this model are not well designed, this has direct bearing on the quality
of the policies that are to be written using that model.
In this paper, we take a software engineering or language-based approach
to designing access policies and present an alternative access control model for
CORBA [16] systems. We believe this model is better suited to express advanced
policy concepts than the default access model specified in the OMG's Security
Service Specification [17]. Also, it provides policy designers with more abstract
concepts for specifying authorizations. This model is based on the concept of
views, an object-oriented approach to specifying access rights.
The rest of this paper is organized as follows. Section 2 outlines the CORBA
access control model and examines the default access control policy as defined in
the Security Service. Section 3 introduces the concept of views and sketches how
it can be used in implicit authorizations and for delegation. Section 4 mentions
related work and Section 5 concludes the paper with a summary and an outlook
on future work.
 238

2 The default CORBA access model

The Security Service specification defines a reference model for access control
which separates access control decisions from the actual access enforcement in
a manner similar to the ISO authorization framework [9]. Access decisions may
be made at both the client and the target side. Within this general framework,
arbitrary access control models may be defined. Note that we use the term access
model to denote a language for defining concrete access control pohcies. The secu-
rity service defines a default access model but calls it the DomainAccessPolicy.
Individual policies are indeed represented by DomainAccessPolicy objects, but
the type DomainAccessPolicy at the same time defines the types of policies that
can be expressed, i.e. the access model. Care should be taken not to confuse the
OMG terminology with the terminology used in this paper.
This section examines the default CORBA access model and identifies some
shortcomings in order to arrive at a set of requirements that a more suitable
access model for CORBA would have to meet.

2.1 Concepts

We briefly explain and review the main concepts of the default access model.

Access matrix. The default authorization model in the Security Service is

based on the access control matrix [14], which represents principals as rows,
objects as columns and permissions or access rights in matrix entries. For a
detailed description of the access matrix semantics of the default CORBA access
control model see [12].
As CORBA is an open distributed architecture, an important requirement
for a general CORBA access control model is scalability. The smallest unit of
access control in object-oriented systems is a single operation on a single object,
but in many cases this level of granularity would induce too much overhead in
operation as well as in management. In order to reduce the number of entities
that need to be dealt with, a number of grouping constructs are provided.

Principals. CORBA refers to both human users and system entities acting on
their behalf a principals. Authorizations are granted to principals only indirectly
on the basis of the security attributes they possess rather than directly using
identities. Possible security attributes include (but are not limited to) access
identifier, group and role names, clearance level and capabilities. If a number
of principals share a common attribute value, they implicitly form a group to
which access rights can be granted.

Domains. Objects are grouped into security policy domains, i.e. sets of ob-
jects to which the same security policies apply. For every domain, there is a
DomainManager object that knows about the pohcies and the members of the
 239

domain. Access policies are represented by a DomainAccessPolicy object by

default.
Domains can be nested, so a domain may contain subdomains and may itself
be a member of an enclosing domain. The main purpose of forming hierarchies of
policy domains is to reflect organizational hierarchies, so subdomains could have
more specific security policies than enclosing domains. The precise meaning of
hierarchically composed policies is, however, not specified. While the implication
seems to be that subdomains have less authority over their members, it is not
clear how conflicts between policies would be resolved.
Another way of composing policies is by forming federations of interoperating
domains, where each domain has full authority but has some amount of trust
in the other and agrees to give certain rights to the other side. As with domain
hierarchies, the question of how this is done and what the semantics of such a
federation would be is left unspecified.

Rights and operations. The security service defines individual rights in rights
families. The default rights family is corba and contains four generic rights: g,
s, m, u for get, set, manage and use. The definition of new rights families, albeit
possible, is discouraged to keep policies simple.
Authorizations are checked per individual operation. The default access model
defines no explicit grouping construct for operations. Operations are, however,
grouped implicitly by the access rights they require. These required rights for
an object access using a particular operation are to be defined by interface de-
velopers and specified per object type. Thus, the specification of required rights
defines a mapping from generic access rights to actual operations. Required rights
are defined as a combination of rights by stating whether the given rights are
required in conjunction ("air'-combinator) or whether any of the listed rights
is sufficient for the operation ("any"-combinator). They are stored in a global
table and are not part of any individual access policy.
Using a domain's DomainAccessPolicy object, a particular access policy
is defined by granting effective rights to principals identified by their security
attributes. One implicit security attribute that is set by the ORB for every
access is the delegation state. This attribute allows to identify for every request
whether the principal is the initiator of the request or an intermediate in a call
chain. It is thus possible to grant a different set of effective rights depending
on the delegation state of the calling principal. Effective rights are registered in
tables and compared to the required rights for an operation upon access. If a
principal's effective rights match the required rights, the access is allowed.

2.2 An example policy

We will sketch a hypothetical access policy for naming context objects using the
CORBA default access model. Naming contexts have the interface CosNaming: :
NamingContext from the OMG's name service specification [15]. The IDL for this
interface is given in Figure 1. The policy is to distinguish between different uses
 240

module CosNaming {
interface NamingContext {
void bindCin Name n, in Object obj);
void rebind(in Name n, in Object obj);
void bind_context(in Name n, in NamingContext nc);
void rebind_context(in Name n, in NamingContext nc);
Object resolve (in Name n ) ;
void unbindCin Name n ) ;
NamingContext new_context();
NamingContext bind_new_context(in Name n) ;
void destroy();
void list (in unsigned long howjnany,
out BindingList bl, out Bindinglterator bi);
};

Fig. 1. Name server interface

of naming contexts with increasing sensitivity. It defines different authorizations

for the following cases:

1. name resolution: only r e s o l v e ( ) operations are allowed. Principals are nei-

ther permitted to learn about new objects by browsing the name context
nor to alter the contents of the name space in any way.
2. additional browsing: allow r e s o l v e ( ) and l i s t ( ) .
3. additional binding of new names: l i s t ( ) , r e s o l v e O and b i n d O are per-
mitted. This case does not include the right to rebind a name, i.e. to bind a
name that is already bound
4. additional context creation: new_context() and bind_new_context() oper-
ations are allowed. These operations return new context objects, the latter
also binds this object to an argument name within the target context, i.e. the
returned context is a subcontext. Note that it is not possible to bind other
contexts as subcontexts than the ones created with bind_aew_context().
The operations bind_context() and rebind_context() are explicitly pro-
hibited to prevent "external" name spaces from being linked into local name
space hierarchies.
5. manage contexts: in addition to allowing all the previous accesses, calling
d e s t r o y 0 is permitted. This will remove a naming context if it is empty.
Also, unbinding and rebinding names is permitted.

Please note that the hierarchical nature of these authorization types is in-
cidental and simply a consequence of the example policy we chose to present
here, it is not characteristic for our approach. While authorizations of the first
two types would be considered " public", authorizations of type 3 and 4 are only
granted to privileged users of two different groups. The last authorization is to
be granted to a small group of administrators only.
 241

Defining Required Rights. Before this pohcy can be expressed in the de-
fault CORBA access model and effective rights can be granted to holders of
the appropriate security attributes, required rights for each operation of the
NamingContext interface have to be defined. The rights combination for each
operation can be specified using rights from
1. the default corba rights family exclusively.
2. another rights family, perhaps one that was introduced for exclusive use with
the type NeimingContext.
3. different rights families.
To keep management of rights simple, it would be desirable to refrain from
defining new rights families and select the first of these options. However, it is
not always straightforward how to model the intended semantics of authoriza-
tion using the generic rights "get", "set", "manage" and "use". Note that these
rights do not just describe distinct operation types (g,s,ii) but also the level of
sensitivity for operations (the m right) [1].
A more serious limitation with using only the small set of rights from the
corba family is that no more than 16 different combinations of rights can be
distinguished. Thus, it is highly Ukely that a number of operations will require
the same set of rights, especially if required rights are defined by individual in-
terface designers and not by security administrators and the domain contains
objects of a number of different types. If required rights are not unique, however,
ensuring the principle of least privilege is very difficult if not impossible with
this access model. Because rights are generic principals can invoke any opera-
tion on any object of any interface in the domain that happens to match the
effective rights combination granted to them. As an example, imagine that the
operation l i s t O on naming contexts and the operation c r e a t e O on some ob-
ject of type JPEGFactory were to require corba :g and corba :u. Any principal
that was granted this combination of required rights in a domain where objects
of these types exist would be allowed both operations and possibly others as
well. Determining which operations will be permitted with a given set of rights
implies searching the whole required rights table because grouping operations by
their required rights can only be done implicitly.
The second of the options outlined above would avoid this problem of in-
terfering rights altogether. Also, defining a new rights family per IDL interface
would allow policy designers to use access rights that directly correspond to the
intended usages of these interfaces. However, introducing new rights families not
only complicates management, it is also very cumbersome in practice as there
are neither language support nor management interfaces for this task.
As a compromise, the basic corba rights could be used in conjunction with a
minimal set of newly introduced, type-specific rights that help ensuring unique-
ness of required rights. For our example, we introduce a new rights family naming
with two rights n, m for naming, and manage. The right naming: n is used to
make this required rights combination type-specific, the second is necessary to
distinguish between the last two authorization types. These are AND-combined
with the default corba rights to specify the required rights as in Figure 2.
 242

required rights combinator operation

corba :—u- all resolve
naming:n-
corba :g-u- all list
naming:n-
corba :gsu- all bind
naming:n-
corba :gsum all new_context
naming:n- bind_new_context
corba :gsum all unbind
naming:nm rebind
destroy

Fig. 2. Required rights

Note that the rights required for, e.g., operation b i n d O include those re-
quired for l i s t 0 . This is to model the policy feature that whoever has authority
to invoke b i n d O can also invoke l i s t ( ) . Also note that to model this property
of the pohcy, we had to give up using the four corba rights as a consistent
classification of operation types: Operation bind() had to require corba:gsu-
although it does not read the state of a naming context, as the presence of the
right g would suggest. The example shows that this access model does not allow
to reconcile the intended semantics with a descriptive and readable specification.

2.3 Limitations of t h e access m o d e l

The default CORBA access model is limited in a number of respects. First,

the default domain access policy does not provide adequate support for very
fine-grained policies. It is not possible to specify different required rights for
individual object instances, nor is it possible to differentiate between domains
by specifying different required rights. If a policy were to require that individual
instances of a type be treated differently, different domains would have to be
set up for each object. Obviously, this will lead to a proliferation of potentially
very small domains when fine-grained access control is necessary, and would
complicate management.
Second, the model will not scale up to systems with large numbers of ob-
jects and interfaces because the authorization requirements expressable using
the CORBA rights concept will frequently coincide. Because rights are used as
generic access types and can practically not be tied to particular IDL types,
enforcing the principle of least privilege is not easily possible within domains.
By granting a set of effective rights, principals will almost invariably gain more
authorizations than they actually require if only the small set of rights defined
in the corba rights family are used — many operations will happen to require
the same rights combination.
 243

Third, the default access model does not define any structural relations for
its access control concepts, e.g. hierarchies of groups, rights or domains. Domain
hierarchies are mentioned in the specification but not defined. Also note that the
management interfaces to the default domain access control model are at a low
level of abstraction and do not offer any language support for the sensitive and
error-prone ta.sks of specifying and managing access rights. Other than domains,
no abstractions are defined that would help structuring large specifications.

3 Views

To remedy the problems pointed out in the previous section we propose a new
access model based on views. A view definition introduces an authorization type
as a named subset of an object type's operations. Thus, individual access rights
directly correspond to operations. Access policies can be described in terms
of principals holding view-typed authorizations on objects. To allow for fine-
grained access control, views can be granted either on a single object of that
type or on all objects in the type's extension that belong to the domain. If a
principal requests access to an object via one of its operations and this operation
is contained in one of the views on that object held by the principal, the access
is allowed.
Note that, although this access model appears to call for a capability-based
implementation, this is not an inherent requirement of the model. 'While individ-
ual authorizations, i.e. view instances, are always related to either a single object
or to an open set of objects (a type extension), they are typed, first-class objects
in their own right. Thus, an implementation of this model is free to represent
the relation between authorizations and objects in any way of its choosing.
Views, i.e. authorization types, help structuring access policies by defining
different usages for object interfaces from the point of view of a security policy
designer. Policy design will obviously benefit from existing use-case diagrams or
role models that describe the intended object interactions from the application
designer's point of view. In spite of this close relation between the design of ob-
ject interfaces and views, it is desirable to keep them separate for two reasons.
First, decoupling interface design from security aspects allows to reuse applica-
tions in contexts with different security requirements or, vice versa, to design a
security policy for existing CORBA applications without having to rewrite their
IDL specifications. Second, it enables us to provide specially tailored abstrac-
tions for describing access pohcies without having to integrate these with other,
orthogonal language features. In particular, we do not want to extend CORBA
IDL with security annotations.

3.1 NamingContext revisited

The authorization types for the example policy from the previous section can be
described directly and conveniently as views as shown in Figure 3.
 244

view CosNamingResolver controls CosNaming: : NamingContext {

allow:
resolve;
};
view CosNamingBrowser : CosNamingResolver {
allow:
list;
};
view CosNamingBinder : CosNamingBrowser {
allow:
bind;
};
vie'W CosNamingContextMeinager : CosNeimingBinder {
allow:
new.context;
bind_new_context;
};
view CosNamingOwner : CosNamingContextManager {
allow:
destroy;
unbind;
rebind;
}:

Fig. 3. Views on name spaces

Views are defined as access controls for a particular IDL interface, which is
referenced in the controls-clause of the view definition. In the example, the base
view CosNamingResolver controls the IDL type CosNaming: :NamingContext.
As a consequence, authorizations granted to give access to objects of one type
cannot interfere with other authorizations because of different types, even if
operation names happened to be identical.
View definitions can be related through inheritance so that their specifica-
tions can be reused. This relation between views is expressed by listing parent
views after the colon. In the example, all view definitions directly or indirectly
extend CosNamingResolver, so they inherit the operation r e s o l v e allowed by
CosNamingResolver and additionally permit the operations listed in their own
definitions. View inheritance is monotonic in the sense that extending views may
only add access permissions.
Apart from syntactic convenience, view definitions have the advantage that
operations are explicitly grouped. It is thus always obvious which operations
will be permitted if a particular view on an object is granted to a principal.
 245

Being able to give names to sets of individual rights also adds descriptiveness as
these names can refer to different usage aspects of a particular IDL type. Note
that views can also be regarded as divisions of IDL interfaces according to the
intended uses of these interfaces, or as retrospectively defined interfaces.
As in the original CORBA access model, authorizations can be granted to
principals indirectly based on the value of some security attribute. The security
service does not mandate any particular way of defining and managing a mapping
of security attributes to users. For syntactic convenience, we suggest to introduce
simple identifiers for predicates over security attributes or combinations thereof
and use these predicate names to denote those principals for which the predicate
holds true.
The view-based access model introduced in this section provides a structured
way of defining rights that is better suited to express complex policies than the
original concept of rights, which was not structurally related to object types. In
the following, we will extend the expressiveness of the view language with con-
cepts that address implicit authorizations and denials as well as the delegation
of access rights.

3.2 Implicit authorizations and denials

An imphcit authorization [18] is one that is imphed by other authorizations.
Authorizations are implied whenever a grouping construct is used for granting.
If, e.g., a view v on the extension of type T is granted to the principals with
the security attribute groupname = staff, this implies granting individual access
permissions for the operations in v to every principal in group staff on every
object in the extension of T.
While it is more convenient to specify general access rules this way than to
grant each of the implied authorizations individually, there also needs to be a way
of expressing exceptions to rules, e.g. that one particular principal in a group is
not to be allowed one particular operation on objects of Type T. Consequently,
it is necessary to define means by which negative authorizations [13] or denials
can be explicitly specified. To this end, we introduce a keyword deny to the view
definition language. Note that no extending view may contain denials because
it can only extend the authorization defined in a parent view.
In some cases it may be necessary for denials to override permissions, but
other cases may demand that permissions can be defined as the exception to a
general rule and thus need to override denials. Hence, a general strategy of re-
solving conflicts between authorizations is called for. A number of strategies are
possible, e.g. "denials take precedence", or bi-valued [18] or multi-level priorities
[4]. Here, we will adopt the distinction between strong and weak authorizations
from [18]: Assuming a partial order is-more-specific on authorizations, a weak
authorization can be overridden by a more specific (strong or weak) authoriza-
tion. A strong authorization, on the other hand, cannot be overridden by any
other, more specific authorization, not even a strong one. The relation is-more-
specific on authorizations is defined in terms of both view hierarchies and object
hierarchies. Principal hierarchies are not explicitly supported in our model but
 246

may be defined ad hoc. They are not considered for resolving authorization con-
flicts.
First, an authorization is more specific than another if it is defined by a view
that is more specific than the view used to define the other, i.e. if the first view
extends the second. An example for such a case is given in Figure 4. Assuming
an IDL definition for a type T with operations op_l to op_5, we may specify a
view BaseView that grants access to operations op_l and op_2 while explicitly
denying op_3 and op_4. Moreover, op_3 is declared as a strong authorization with
the keyword f i n a l , so it cannot be overridden in refining views. A view definition
that attempted to do so would be rejected. DerivedView refines BaseView and
is thus more specific, so its authorizations will override the ones inherited from
BaseView in case of conflicts. Such a conflict arises because DerivedView allows
op_4, which was inherited from BaseView as denied. Additionally, DerivedView
allows to access objects of type T using op_5.

view BaseView controls T {

allow:
op-i;
op-2;
deny:
final op_3;
op_4;
}:

view DerivedView : BaseView {

allow:
op-4;
op-5;
};

Fig. 4. Weak vs. strong authorizations

Second, an authorization ai is considered more specific than another autho-

rization 02 if a2 is granted on the extension of some type T and ai is granted
on either the extension of a subtype S of T, one particular instance of S or a
single instance of T. As is-more-specific is partial, there are cases where both
the views and the involved objects of ai and a2 are unrelated. The relation is
also not defined if ai would be more specific than a2 because of its more derived
view while at the same time 02 would be more specific than Ci because of the
type structure of the involved objects. In these cases, denials take precedence.
To decide whether an access request can be allowed it is necessary to search
the initiator's views for both permissions and denials. The access is allowed
if a permission but no denial for the operation is found. If both a permission
and a denial are found, the access decision depends on the structural relation
between the authorizations. If the permitting authorization is more specific than
 247

the denying authorization, the access is allowed. If the denying authorization is

more specific or the authorizations are unrelated, the access is denied.

3.3 Delegation of authorizations

Some situations may require principals to delegate some or all of their own rights
to other principals. This can be necessary, e.g., for cooperation purposes, for del-
egating authority or as part of other application protocols. The CORBA security
reference model describes a number of different privilege delegation schemes to
support call chains, i.e. delegation is necessary because calls may be delegated.
Principals pass their credentials with the access request. In the no delega-
tion scheme, the target object uses these credentials for access checks but is not
allowed to use them for accessing other objects. If calls are delegated and the
target becomes an intermediate in a call chain, it has to use its own credentials
(if any). In simple delegation — also called impersonation — the intermediate
may use the received credentials in subsequent calls and may also delegate these
further. Only these two schemes are required for compliant security service im-
plementations. Two more delegation schemes, composite delegation and traced
delegation are possible.
The effective delegation scheme to be used is selected by an administrator.
This decision is taken for a delegation policy domain which need not be identi-
cal with the access policy domain. Security-aware applications can also set the
delegation scheme according to application policy. The security service mentions
but does not specify interfaces for achieving controlled delegation by limiting
the vaUdity period of delegated credentials or the number of invocations they
may be used for. Restrictions on the target objects are also envisioned, but not
currently defined.
In principle, this credentials-based delegation approach does not rule out
implementations that pass authorizations directly, using capabilities [6] or au-
thorization certificates, e.g. SPKI [5] or X.509 certificates. However, the CORBA
delegation schemes actually aim at delegating security attributes (as indicated by
calling simple delegation impersonation) and thus transfer authorizations only
indirectly by changing the set of principals that possess these security attributes.
As a consequence, it is difficult for a principal to infer and control which
of its own authorizations are actually delegated. Also, authorizations can only
be delegated implicitly by calling other objects — there is no explicit "grant"
operation that would simply transfer some authority to other principals. As a
last observation, note that it is not possible to specify restrictions on delegation
in an access control policy as delegating security attributes does not change the
access rules defined in the policy.
For these reasons, we propose to add the meta-operation "grant" to our ac-
cess model and mark view definitions as grantable if delegation of authorizations
of this type is permitted. Delegating views on objects is an explicit operation,
so applications wishing to delegate rights need to be security-aware. Note that
there is no "revoke" right. Granting an authorization implies the right to revoke
it any time.
 248

A delegating principal might wish to restrict the delegated rights in a number

of ways. First, it might wajit to ensure that the recipient may only invoke those
operations that are necessary for a particular task. Second, it might want to
disallow further delegation of its rights to other principals.
To allow controlled delegation of access rights, we introduce an additional
grant operation for views. A view that lists this operation as allowed permits
a principal to delegate authorizations of this type. If it does not allow grauit,
delegation is not permitted. The grant operation can optionally be parameterized
with identifiers referring to principals or principal groups, thus restricting the
range of recipients for a view to a specific subset of principals. An example of
such a view definition is given in Figure 5.

view DelegatableView controls M::T {

allow^:
grant ( t e c h - s t a f f . r i c h a r d s ) ;
op-1;
deny:
op-4;
};

Fig. 5. View delegation

In this example, a principal holding a DelegatableView authorization can

delegate it only to the principals that are associated with either the identifier
tech_staf f or r i c h a r d s . If principals wish to impose further restrictions when
delegating views, they can dynamically derive new, more limited authorizations
from the ones they hold. Deriving new authorizations may only be done in a
restricted manner by selecting a parent view as the type of the derived autho-
rization. To prevent denial of service attacks, only authorizations without denials
may be delegated, so delegation can only increase a receiving principal's access
rights.

3.4 Combining access rights

Our access control model unifies the concepts of operation and access right: View
definitions list operations, and holding a particular view on an object permits
to access this object with precisely the operations listed in that view. Since
operations are indivisible, so are access rights. While we believe this model to be
more appropriate for object-oriented systems than approaches which separate
access rights from operations, our model lacks the flexibility offered by the option
of combining access rights to obtain a certain privilege.
As an example, consider an access control policy for a safe with multiple locks
that states the safe can only be opened when a certain number of keyholders
cooperate. Accessing the safe should require multiple rights which correspond to
the keyholders' keys. Using the model outlined so far, however, we cannot express
 249

conditions on operations that require the caller to possess further authorizations

than just the one that permits the operation.

To provide a flexible way of expressing access rights that depend on a combi-

nation of other rights, we introduce a new keyword requires as a modifier for a
simple access right. In the example in Figure 6, holding a SecuredSaf eView on
an object is not sufficient to be allowed to access that object using open. Prin-
cipals that want to do so need to hold other views that allow k e y l , key2 and
key3, such as KeyholderOneView. This view allows keyl but lists it as v i r t u a l .
By declaring an authorization as virtual, we allow to decouple actual operations
in interfaces and access rights: a virtual right has no corresponding operation
in an IDL interface and thus does not allow any direct object access. The sole
purpose of having virtual authorizations is to avoid having to introduce unnec-
essary operations into IDL interfaces simply because combined access rights are
needed.

view SecuredSaf eView controls M: :Safe {

allow:
close;
open r e q u i r e s KeyholderOneView.keyl,
KeyholderTwoView.key2,
KeyholderThreeView.key3;
};

view KeyholderOneView controls M::Safe {

allow:
gr£int{keyholder};
virtual keyl;
};

Fig. 6. Combined access rights

When a principal tries to open a Safe object, the access decision function
checks not only that the principal holds a view that lists open as allowed, but
also whether the principal possesses the additional authorizations k e y l , key2
and key3 for this object that are required by open. If it does not, the principal's
access will be denied and it will have to obtain the missing rights, e.g. by having
other principals grant them. In cases like this, where the cooperation between
principals requires the exchange of access rights for a single operation invoca-
tion, the introduction of further delegation restrictions appears necessary so that
granting principals could specify how often access rights may be used or for how
long a granted view remains valid.
 250

4 Related work

Using type information for specifying fine-grained object protection is, of course,
not new. Language based-approaches such as [11,19], typically rely on an ex-
tended notion of type that allows to selectively hide some of an abstract data
type's operations for access control. We are not, however, aware of approaches
that establish a separate type-concept for authorizations.
Views are a well-known concept in relational and object-oriented databases
[20]. Their use for access control purposes resembles the use of type abstraction
as a protection concept. Unlike database views that can span multiple types, a
view in our model is restricted to objects of a single IDL type. Joining views
on different IDL types Ti,...,T„ can, however, be modelled by specifying an
additional IDL interface T that extends Ti,...,T„ and defining a view on T.
Another difference is that database views may define content-specific access
controls, e.g. by stating that an attribute may only be read if its value is above
a certain threshold. This is not possible in our model.
In [10], grouping concepts for objects, principals and access rights have been
proposed in the context of downloadable code, but without proposing a language
for the specification of access control pohcies. While this model also addresses
implicit and negative authorizations, rights delegation and combination are not
an issue.
In [8], views are lists of allowed operations that are used to define the access
rights contained in capabilities. By recursively annotating parameters of view
operations with other view names in interface definitions, protection require-
ments for arguments can be expressed. At run-time, access rights according to
these views are passed implicitly. Structural relations between views and nega-
tive rights are not addressed.

5 S u m m a r y and future work

We have analyzed the CORBA default access model and identified a number
of restrictions that make it desirable to replace this access model. We have
presented and dicussed views as the basic concept of a new access control model
that will allow policies to be expressed at a higher level of abstraction. We have
also sketched how implicit authorizations and delegation could be expressed in
access policies based on this model.
We are currently working on an implementation of the concepts presented in
this paper as part of a partial security service implementation for our own Java
implementation of the CORBA specification, JacORB [3]. In this implementa-
tion, view definitions are compiled into runtime representations that reside in
a view repository. The function of such a view repository is similar to that of
the CORBA interface repository. We are currently evaluating the possibility of
representing authorizations in the form of SPKI certificates [5]. Future work in-
cludes formalizing the concepts presented in this paper and exploring how role
models could be included.
 251

Another question in this context that we intend to pursue is how interoper-

ability between security policy domains can be supported. The security service
specification does mention both hierarchies and federations of domains, but does
not answer the question of how these federations are defined and what proper-
ties a resulting, composite policy would have to have. Moreover, in an open
distributed architecture such as CORBA, it might not be possible to pre-define
federations of policies for all possible situations, so it will be necessary to be able
to establish interoperation dynamically in some cases.
In order to allow cross-domain accesses at all, a domain must be able to
specify how much it trusts external principals and on what properties this trust
is based. This determines what authorizations it is willing to grant to principals.
Taking a pragmatic approach, the amount of trust a domain is prepared to invest
in a principal can be directly identified with the authorizations it grants as in
PoUcyMaker [2] or SPKI [5].
This does not, however, address the issue on which grounds trust or autho-
rization is granted. Granting certain authorizations to, e.g., a public key amounts
to a certain level of trust being placed in the holder of that key uncondition-
ally, simply on the basis of the existence of the key. Any supposed properties
or characteristics of the entity holding that key are not expressed. One simple
and straightforward way of specifying on which attributes of a principal trust
is based is to grant authorizations to principals indirectly, i.e. on the basis of
certain security attributes as mentioned in this paper. Security then depends on
proving these properties, i.e. being able to do cross-domain authentication. The
practical problem here is to define a mapping between external and internal se-
curity attributes. Other, theoretical problems include reasoning about properties
of composite policies (cf. [7]).

Acknowledgements

I would like to thank Peter Lohr and Richard Kemmerer for valuable discussions.

References

1. Blakley, B., http://www.bhs.org/IT/Projects/cpr/security/CORBASEC-FAQ/,

Section 4.3.2, April 1998.
2. Blaze, M., Feigenbaum, J., Strauss, M.: Distributed Trust Management. Proc.
IEEE Symposium on Security and Privacy (1997) 143-154
3. Brose, G.: JacORB — design and implementation of a Java ORB. Proc. Dis-
tributed Applications and Interoperable Systems DAIS'97, Chapman &; Hall
(1997) 143-154
http://www.inf.fu-berlin.de/"brose/j acorb/
4. Briiggemann, H. H.: Spezifikation von objektorientierten Rechten. DUD-
Fachbeitrage. Vieweg (1997)
5. Ellison, C, Prantz, B., Lampson, B., Rivest, R., Thomas, B., Ylonen, T.: SPKI
Certificate Theory. Internet Draft, draft-ietf-spki-cert-theory-02.txt (1998)
 252

6. Gong, L.: A secure identity-based capability system. Proc. IEEE Symposium on

Security and Privacy (1989) 156-83
7. Gong, L., Qian, X.: Computational issues in secure interoperation. IEEE Trans-
actions on Software Engineering, Vol. 22(1), January 1996, 43-52
8. Hagimont, D., Huet, O., Mossiere, J.: A protection scheme for a CORBA envi-
ronment. Proc. ECOOP '97 Workshop, CORBA: Implementation, Use and Eval-
uation (1997)
9. ISO/IEC: Information Technology — Open Systems Interconnection — Security
Frameworks for Open Systems: Access Control Framework. ISO (1996)
10. Jaeger, T., Rubin, A., Prakash, A.: Building systems that flexibly control down-
loaded executable content. Proc. 6th USENIX Security Symposium (1996) 131-
148
11. Jones, A., Liskov, B.: A language extension for expressing constraints on data
access. Communications of the ACM, Vol. 21(5), May 1978, 358-367
12. Karjoth, G.: Authorization in CORBA Security. Proc. ESORICS'98, LNCS 1485,
Springer (1998) 143-158
13. Kowalski, O., Hartig, H.: Protection in the BirUX operating system. Proc. 10th
Int. Conf. on Distributed Computing Systems (1990) 160^166
14. Lampson, B.: Protection. ACM Operating Systems Review, Vol. 8(1), January
1974, 18-24
15. OMG: CORBAservices: Common Object Services Specification. OMG document
formal/97-12-02, Object Management Group (1997)
16. OMG: The Common Object Request Broker: Architecture and Specification, re-
vision 2.2. Object Management Group (1998)
17. OMG: Security Service Revision 1.2. Object Management Group (1998)
18. Rabitti, F., Bertino, E., Kim, W., Woelk, D.: A model of authorization for next-
generation database systems. ACM Transactions on Database Systems, Vol. 16(1),
March 1991, 88-131
19. Richcirdson, J., Schwarz, P., Cabrera, L.-F.: CACL: Efficient fine-grained protec-
tion for objects, Proc. OOPSLA 1992, 263-275
20. Scholl, M., Laasch, C , Tresch, M.: Updatable views in object-oriented databases.
Proc. 2. Int. Conf. on Deductive and Object-Oriented Databases, LNCS 566,
Springer (1991) 189-207
 Apoptosis - the Programmed Death
of Distributed Services

Christian Tschudin

Abstract. Active networks enable to deploy new services at run-time by us-

ing mobile code. While considerable effort is under way to build active network
infrastructures and to understand how to create corresponding services, less is
known about how to end them. A particular problem is the coordinated steering
of mobile code based services, especially in the case of "strong" active networks
where each data packet is replaced by a mobile program and where a distributed
service can consists of a myriad of anonymous active packets. In this paper we
introduce the concept of apoptosis for mobile code based services. This term is
borrowed from cell biology and designates the programmed cell death. We dis-
cuss the need for a self-destruction mechanism inside a distributed mobile service
and address the problem of securing such a mechanism against malicious activa-
tion, for which a simple solution is shown.
Keywords: Active networks, mobile code security, apoptosis.

1 From Mobile Servers to Highly Distributed Services

Most network services today are provided by stationary programs, either at the appli-
cation level (e-mail) or at the network level (routing). Programmable networks enable
to reconfigure the network's nodes and to bind servers to new physical locations at
run-time. "Network-aware services" may choose different server locations for optimiz-
ing the quality of service [5]. Similarly, application level gateways were proposed that
can perform transcoding or downgrading of multimedia data [1]. Within such proxy
architectures, the thin clients - typically mobile devices with wireless links to the fixed
network - can program the gateway by uploading servlets. In both cases we have rather
large servers and services with a limited amount of mobility.
By reducing the granularity of mobility, we can think of distributed network services
that consist of many tiny mobile programs going forth and back. Maintaining routing
tables, for example, does not require large server programs sitting somewhere in the
network: Small active packets can collect and propagate topology changes to all nodes,
updating the routing tables directly. A natural form of mobility of such services consists
in having them look around for "unserved" nodes to which the service will attempt to
extend. Such services will self-deploy and form a floating, gas-like service cloud that
reaches into every niche of the network [ 12]. It is also possible that such a service cloud
retracts from selected nodes because offering the service is not viable anymore for some
places due to increased resource competition or lack of clients. For the rest of this paper
we will focus on such fine-granular mobile code based network services and call them
highly distributed mobile services (HDMS).
 254

1.1 Steering Distributed Mobile Services

In current distribute systems it is the duty of human operators to steer a service. This also
includes the update of software because of changes in the protocol standards or bug fixes
in the implementation. While this does not happen very frequently with todays large and
non-mobile servers, we expect this to be a fairly common case in active networks simply
because it is so easy to deploy new functionality. The question then is how to steer such
a mobile service in an active network. One of the central steering functionalities will be
the termination of a distributed service.
While it is possible to track the heavy weight servers and to establish special control
channels to them, such an approach is useless in the HDMS case. The sheer number of
tiny elements executing in the network, and the network's topology changes leading
to a never-ending adaption process of HDMS services, makes it impossible to locate
and address the single elements in a centralized way. Terminating a service in this case
will rather be a question of using another HDMS service that will "spread the word".
Because the targeted HDMS service knows best how and in which niches it extends, it
is useful to build a self-destruction capability into the service. Once triggered by some
external event, a termination signal will propagate along the same line as the HDMS
service deployed itself. This enables to shut down a HDMS service in an ordered way
such that we can speak about the programmed death of a mobile service. The overall
effect is a joint service suicide.
The concept of programmed death is known in biology as apoptosis. This word,
pronounced either as APE-oh-TOE-sls or uh-POP-tuh-sis, is Greek and stands for "to
fall off". In section 2 we first describe apoptosis from a cell biology's point of view.
In section 3 we link these observations to the problems of protecting the termination of
HDMS services and present a simple way to weld together a specific death signal with
the self-destruct routine. We close this paper by a brief discussion of related work and
with an outlook.

2 Apoptosis
Research in biology has revealed that cells have a limited capacity to divide (mitosis).
This is not due to physical limitations - like for example exploiting some resource
beyond usability - but is a predetermined, intrinsic behavior of the cell. Mechanisms at
the molecular level are in place that can trigger the self-destruction of a cell. Several
reasons have been identified why it makes sense that a cell commits suicide (two recent
publications that give an overview of this field are [9,2]);
Cell death is as important as mitosis:
During the growth of an organism and the specialization of cells it is necessary
that some cells yield the place they occupy. Fingers, for example, are formed by
apoptosis of the tissue between them. Another example is the formation of the con-
nections between neurons in the brain that requires that surplus cells be eliminated
by apoptosis.
Combating ceils infected with a virus:
Cytotoxic T lymphocytes (CTL) can kill virus-infected cells by inducing apoptosis
i.e., killing the cell and the virus.
 255

Steering the response of the immune system:

The immune system responds to a viral attack by creating CTLs. In order to reduce
the immune response after the attack, CTLs induce apoptosis in each other and even
in themselves. Defects in this machinery are associated with so called auto-immune
diseases where CTLs continue their work and attack body constituents.
Elimination of cells with damaged DNA:
DNA damage leads to an increased cell-internal production of the p53 protein,
which is a inducer of apoptosis. Apoptosis helps to eliminate cells with broken
DNA before it can lead to birth defects. This process also aims at preventing that a
cell becomes cancerous. In fact, many cancer cells have a mutation in the gene that
encodes the p53 protein!

2.1 Apoptosis vs. Necrosis

The important aspect of apoptosis is that the cell's self-destruction proceeds in a pro-
grammed and controlled way. The cell starts to shrink, it decomposes internal structures
and degrades all internal proteins. Then the cell breaks into small, membrane-wrapped
fragments ("fall off") that later on will be engulfed by phagocytic cells for recycling (see
figure 1). This process does not induce inflammation, no toxic substances are leaked to
the cell's environment.

Fig. 1. Two forms of cell death: Programmed cell death (apoptosis) and necrosis due to injury.

The other form of cell death is necrosis, the wn-programmed death of a cell. This
happens if a cell is injured, either by mechanical damage or exposure to toxic chemicals.
The response will be that the cell swells because the membrane's function is disrupted.
 256

The cell contents leaks out which eventually leads to an inflammation of the surrounding
tissue until phagocytic cells removed all cell debris.
In cell biology there are two ways known how to induce apoptosis. A cell can be led
to commit suicide either by withdrawing specific positive signals i.e., signals needed for
continued survival like grow factors. Or, apoptosis is triggered by negative signals like
increased level of oxidants within the cell, UV light, X-rays, chemotherapeutic drugs as
well as death activators specifically installed to induce apoptosis.
Surprisingly enough, the protection against an "unauthorized" triggering of a cell's
self-destruction mechanism is not the major issue in biology. Because viruses as well as
cancer genes require a cell to function for their own procreation, they will rather try to
avoid destroying their host. This means that there is a natural selection of benign genes
that are able to block apoptosis procedures. The battle between viruses and the immune
system thus is rather a question of keeping the apoptosis mechanism intact.
In the context of active networks we have a different constellation. Active packets
in general do not vitally depend on each other but execute independently of each other.
Instead of a parasitic relationship there is competition between them, making the elim-
ination of a rival mobile service a valuable goal: The apoptosis entry point of a mobile
service would be an primary target for an attack. In the following section we will show
a simple way to protect this deadly entry point.

3 Apoptosis of Distributed Mobile Services

Similar to the case of a cell, we can identify important reasons why we must be able to
let a mobile service self-destruct:
- The service has reached a state from where it does not know how to return to a
safe configuration. This may be due to a runtime error, or due to the discovery of
corrupted and compromised data.
- The service feels it is under attack and prefers to destroy the service and its secrets
altogether.
- A new version of a running service should be installed that requires the old one to
completely give way.
In analogy to the biological case there is a benefit in doing a service shutdown in an
ordered way. The programmed and controlled termination helps to start a follow-up
service by letting it proceed from a known state, without being fooled by residual data
and code traces. Furthermore, lingering fragments may be subject to some economic
charging scheme inside the network or would uselessly intensify the competition for
the network's scarce resources.
As for cells we can see two different ways for a distributed service to start the
apoptosis process:
Absence of positive signals:
A mobile service may depend on a continuous stream of credentials. For example,
in an active network using economy inspired resource management, these signals
would be a form of electronic money. Once a service realizes that it will soon run
out of money, it can shut down partially or even completely.
 257

Presence of negative signals:

A service that discovers modifications of its code base will probably decide to stop
immediately. Some hardware systems even have this behavior built into them (e.g.,
the PC's memory parity check). We must also consider external sources of a nega-
tive signal, for example a network manager that wants to terminate the service that
he created.
From a security point of view, positive signals are vulnerable to denial of service
attacks. By the mere suppression of positive signals it is possible to bring a mobile code
based service down, if it bases its apoptosis decision on the absence of positive signals.
There is little that can be done against denial of service attacks at the level of active
packets or with resource manipulations by a malicious host. We will not consider this
case further in this paper.
Negative signals, on the other hand, have the problem that they can be faked. The
executing host is in a particularly powerful position because it has access to the service's
code and can figure out the decision logic that starts the "tear down" procedure as well
as this procedure's code. But even in the case of a trustworthy host there is a need to
protect the "negative signal" channel. Otherwise, attacks can be mounted by other nodes
or mobile services that know the format of a negative signal message.
In the following we will look at the second problem. We show a code transformation
technique that enables to securely use negative signals in an unsecured context such that
arbitrary parties can try to induce the apoptosis and where malicious hosts may peek
and poke the executing image of a mobile service at will.

3.1 Securing the Apoptosis Activator

In the context of active networks the "apoptosis activator" translates to the name and
code of a self-destruction routine. A specific death signal shall be defined that activates
this routine which starts the termination process, including the propagation of the signal
to other parts of the distributed service. To this end, each element of a service has to
provide its apoptosis activator. Three items need to be protected:
1. The activation signal:
otherwise anybody can create a false apoptosis signal. The signal thus has to be
encrypted or otherwise hidden.
2. The decision logic:
otherwise a malicious host can override the conditional execution and make it un-
conditional. The apoptosis trigger code thus needs execution integrity.
3. The self-destruction routine:
otherwise a malicious host can immediately access and execute the apoptosis rou-
tine. The code thus has to be encrypted.
We want to protect the following prototypical piece of code:
s i g n a l := r e a d ( ) ;
IF signal = secret_signal THEN
self-destruct();
FI
 258

We see that the execution of the self-destruction routine depends on "environmental

data" requested by the r e a d () instruction. How this data is gathered is usually outside
the control of a mobile program. The only thing one can hope to secure is the way this
data is processed. What we would like to do is to make this mobile code fragment a
"clue-less mobile agent" in the terminology of Riordan and Schneier [7]. Although the
agent's actions depend on data provided by the environment, it should not be be possible
to infer which environmental value will trigger what action.
This goal can be achieved by transforming the previous code fragment into the
following "wrapper" program:

key := read() ;
result := decrypt(key, ENCRyPTED_CODE);
IF is_valid(result) THEN
execute(result);
FI

The ENCRYPTED_CODE constant is computed at the originator's site by encrypting the

self-destruction code with the chosen signal. This constant byte array is then wrapped
into the small program shown above. Note that neither the unencrypted code nor the
signal's value appear in this program.
At the executing site, the now secured mobile program waits for a signal to be pre-
sented as environmental data. The received value becomes a key that the mobile pro-
gram attempts to use for decryption. For each presented key we attempt to decrypt the
ENCRYPTED-CODE. Looking at the result we can decide if this was a valid decryption
or not.
Note that this final sanity test can be omitted if we do not care about having the host
executing a nonsense program. The test, however, is simple and does not add any addi-
tional burden to the process. It suffices to include a message digest inside the encrypted
code for comparison purposes.

3.2 An Implementation

The simple code transformation presented above was implemented in July 1997 for the
MO system [11]. MO lends itself very well for this because of its interpreted nature
and the easy with which code can be generated at runtime. A simple procedure was
written that takes a pass-phrase together with the MO code that is to be secured. The
pass-phrase is hashed using the MD5 digest algorithm, the lower 64 bits are used as
the DES key for encryption. The code, together with a nonce value and a message
digest of the code, is then encrypted and emitted inside the "wrapper" code as described
in section 3.1. The wrapper consists in a MO procedure that takes an environmental
data item, applies the hash function, decrypts the given string constant, checks for the
correct digest value and, if correct, requests the execution of the decrypted code. The
transformation procedure itself fits in 12 lines of MO code.
 259

4 Discussion

The presented welding together of an equality test with the conditional code is described
in [7], as are further possible code transformations that enhance the "ciuelessness" of
mobile code. Note that the presented approach works for terminating a distributed ser-
vice because we do not require the privacy of the apoptosis code to be maintained be-
yond the moment where we want the service to terminate. To obfuscate or completely
hide the functionality of a piece of executable code for longer time spans is a much
harder problem [3,4,8]. Our hope is that more simple building blocks as the one pre-
sented in this paper will be discovered, making security sensitive software fully mobile
and deployable in unsecured environments. Active networks will be an important appli-
cation domain for this.

4.1 Artificial Death

The programmed death topos is already present in computer science, namely in the
domain of artificial life. These "alife" simulations have to deal with the problem of
damping the procreation. For example, genetic algorithms have a selection step that
imposes a limit on the individuals that can proceed to the next evaluation round. This
leads to the "artiiicial death" of the less fit individuals. Or, in the Tierra system [6], an
explicit reaper function eliminates the individual that made the most mistakes, while
other artificial life systems introduced a suicide instruction. This topic is discussed in
more general terms in [10]. Although artificial life experiments were carried out mostly
in a non-distributed setting, today they start to enter the networked environment via
distributed games.

5 Conclusions: Merging Control and Data

The analogy proposed in this paper between the programmed death of a cell and the
self-termination of a distributed service points to the question on how active networks
will be steered in the future. From an engineering point of view it makes sense that
if we are required to firmly control the execution of mobile services, then we should
built the support for this into the active network substrate. Thus, the approach would
be to create privileged control channels that are different (or even separated) from the
rest of the network infrastructure. However, it may be difficult, if not impossible, to
come up with a universally applicable, reliable and scalable termination service. As an
indication of the problems to be expected one can look at the related problem of finding
the reliable broadcast protocol, or the effective support for distributed debugging. In our
research we will continue to handle control and data at the same level, as the insights
from biology suggest.
 260

References
1. Amir, E., McCanne, S. and Katz, R.: The Media Gateway Architecture: A Prototype for
Active Services. Proc. SIGCOMM'98, Vancouver, Canada, 1998.
2. Bowen, I.D., Bowen, S.M. and Jones, A.H.: Mitosis and Apoptosis - Matters of Life and
Death. Chapman & Hail, 1998.
3. Collberg, C , Thomborson, C. and Low, D.: Manufacturing Cheap, Resilient, and Stealthy
Opaque Constructs. Proc. Principles of Programming Languages 1998 (POPL'98), San
Diego, California, Jan 1998.
4. Hohl, R: Time Limited Blackbox Security - Protecting Mobile Agents From Malicious
Hosts. In Vigna, G. (Ed.): Mobile Agents and Security. LNCS 1419, Springer, April 1998.
5. Ranganathan, M., Acharya, A., Sharma, S., Saltz, J.: Network-Aware Mobile Programs.
Proc. USENIX 97, Anaheim, California, USA, 1997.
6. Ray, T: An Approach to the Synthesis of Life. In Langton, C , Taylor, C , Farmer, J. and
Rasmussen, S. (Eds): Artificial Life II, Redwood City, CA, 1991.
7. Riordan, J. and Schneier, B.: Environmental Key Generation Towards Clueless Agents. In Vi-
gna, G. (Ed.): Mobile Agents and Security. LNCS 1419, Springer, April 1998.
8. Sander, T. and Tschudin, C: Towards Mobile Cryptography. Proc. IEEE Symposium on
Security and Privacy, Oakland, May 1998.
9. Sluyser, M. (Ed): Apoptosis in Normal Development and Cancer. Taylor & Francis, London,
1996.
10. Todd, M.: Artificial Death. Proc. 2nd European Conference on Artificial Life (ECAL93),
Brussels, Belgium, 1993.
11. Tschudin, C: The Messenger Environment MO - a Condensed Description. In Vitek, J. and
Tschudin, C. (Eds): Mobile Object Systems - Towards the Programmable Internet. LNCS
1222, Springer, April 1997.
12. Tschudin, C : A Self-Deploying Election Service for Active Networks. To appear in Proc.
3rd Int. Conference on Coordination Models and Languages, Amsterdam, April 1999.
LNCS, Springer.
 A Sanctuary for Mobile Agents

Bennet S. Yee

1 Introduction
The Sanctuary project at UCSD is building a secure infrastructure for mobile
agents, and examining the fundamental security limits of such an infrastructure.
First, what do we mean by "secure"?
An obvious issue is the privacy of computation. With standard approaches
for agent-based systems, a malicious server has access to the complete internal
state of an agent: software agents have no hopes of keeping cryptographic keys
secret in a realistic, efficient setting. Distributed function evaluation approaches
may seem to apply, but that requires an unrealistic fault model and is not likely
to ever be practical. Approaches such as [1,2,27] are extremely expensive or
have very restricted domains.
The privacy of computation is only one aspect of the security picture: the in-
tegrity of computation is perhaps more critical. In agent-based computing, most
researchers have been concentrating on one side of the security issue: protecting
the server from potentially malicious agents. Related work in downloadable ex-
ecutable content (Java [13], Software Fault Isolation [29], Proof-Carrying Code
[24,25], OS extension mechanisms such as packet filters [21], type safe languages
[9,16], etc) all focus on this problem. The converse side of the agent security
problem, however, is largely neglected and needs to be addressed: how do we
protect agents from potentially malicious servers? Why should we beheve that
the result returned by our software agents are actually correct and have not been
tampered with?

2 Software Agents and Malicious Servers

In agent-based computing, not only do servers fear that agents bring in viruses
or attempt to subvert the server, but the agent's user also needs to be able to
trust that the agent was not subverted when visiting a series of servers, some of
which may be malicious.
A simple example of how such a subversion might occur will make this prob-
lem clearer. Let's look at the standard air-fare agent scenario: I need to travel to
Washington D.C. to attend a meeting, so I send a software agent to visit servers
at all the airlines to query their databases to determine the least expensive air-
fare from San Diego to Washington D.C, subject to various trip timing, seat
preference, and routing constraints. One of the airlines, Fly-By-Night Airlines,
runs a server, www. f lybynight. com, where my agent's code is automatically rec-
ognized and "brainwashed": its memory of what other airlines it has visited and
what prices it had seen is modified, so that it ends up recommending a "red-eye"
 262

flight by Fly-By-Night Airlines, when a less expensive daytime flight oflFered by

another airline would really have been preferred.

3 Partial Solutions
How can software agents be protected from malicious servers? This is a critical
security problem to be solved if we are to have faith in agent-based computing.
In the following sections, we will examine several approaches and discuss their
limitations.

3.1 Legal Protection

One approach to the agent security problem is via legal/contractual means.

Operators of the servers where agents run promise, via contractual guarantees,
that they will keep their servers secure from external attackers and that they
will not violate the privacy or integrity of the software agents' computation. No
complex cryptographic protocols are required — there are no run-time overhead
at all!
Such an approach, however, is not entirely satisfactory: for it to work the abil-
ity to detect breaches of contract is still critical. Furthermore, for that detection
to be meaningful, tamper-proof logs must be available to serve as non-repudiable
evidence of the breach of contract should lawsuits become necessary.

3.2 N o Protection

For certain classes of computations, no protection is necessary, and if we are to

carefully examine the cost/benefits of providing protection for software agents,
this must be examined. What types of computations require no protection? Sup-
pose the result of the computation is easily verifiable, e.g., the existence of an
airfare that is below $200. In this scenario, agents may simply replicate and
flood-fill all airline servers to make sure that a copy of the agent has run on each
server, and each agent copy can send the corresponding flight information if it
finds one that costs less than $200. No agent state needs to be transferred at all.

3.3 Fault Tolerance Approaches

In this section, we discuss fault-tolerance-style approaches to the agent security

problem. First, we make some general observations on which aspects of agent
state are vulnerable to attack, and which aspects may be systematically verified.

Observations First, note that uncorrupted servers can determine whether

agent code and read-only state have been modified: the originator of the software
agent can digitally sign the agent code and all read-only configuration variables
before dispatching the agent to agent servers, and the agent servers can verify
 263

both the origin and the integrity of these aspects of the agent. (Message au-
thentication codes are inappropriate, since potentially malicious agent servers
should not share secrets with the originator of the agent.) Other than crypto-
graphic techniques (if any) needed for the secure communication links, for now
we will not require the servers to perform any cryptography.
It may seem that agent code signing could be circumvented by a malicious
server, since the malicious server could tamper with the agent and then re-sign
it with its own key. This approach, however, is thwarted by the following design:
agents are constrained to send its results only to the entity that signed them.
Thus, conceptually a server that re-signs an existing agent is simply performing
two actions at once: denying service to the true originator of the agent, and
sending out its own agent, possibly with initial data stolen from the "murdered"
agent.
Next, note that the originator can specify the order in which the software
agent will visit the airline servers. Abstractly, this is a circuit of the (complete)
graph connecting the airline servers, and the originator may chose this circuit at
the time of agent dispatch.
At any honest server, the agent code and its read-only state is checked when
the agent arrives, so if the malicious server tries to tamper with the agent code
or the read-only state the malicious server can not successfully pass the modi-
fied agent to an honest server. (Alternatively, the agent code and the read-only
state may be considered to be reloaded from the originator by every server.) Fur-
thermore, we assume that the variable agent state is transmitted among servers
using authenticated and encrypted channels, so that only the server that is the
intended migration target can receive the agent, as long as the agent is starting
from an honest server. Thus, the malicious server can not intercept an agent as
it migrates from an honest server to another server.
At any server, an agent may query the server's identity. At first glance, this
identity could be authenticated via a public key certification chain, with the root
certificate embedded as part of the read-only agent state. Note, however, that
the use of cryptographic authentication does not really help a software agent
to determine the hosting server's identity: since the server has control over the
agent's computation, the mahcious server may simply cause the program counter
to bypass the cryptography-based identity query and force the program to take
the conditional branch(es) which corresponds to the desired (falsified) server
identity.
In addition to being able to ask for the identity of the current server, the
agent may also ask from which server did the agent migrate. Because we assumed
that server-to-server communications use authenticated and encrypted channels,
servers will know from which server did an agent arrive. If the agent is running on
an honest server, both these answers will be correct and they can be used to verify
that the agent had migrated on an edge on the intended migration circuit; if the
agent is running on the malicious server, these answers may be incorrect and
the agent's state may be modified so that it believes it is running on a different
server. In the special case where there is exactly one malicious server, this ruse
 264

will be discovered when the agent migrates off of this server to an honest server.
If there are two or more malicious servers, the first malicious server encountered
by an agent can hand it off to any of the other malicious servers in the route
the agent is programmed to take. When the agent is passed on to the next
(honest) server, the agent is brainwashed to believe it had visited all the servers
in the original path between the two malicious servers, thus avoiding discovery. / /
software agents are to depart from the route determined at agent dispatch time,
such departures must start and end at a malicious server.
Now, consider what visiting a malicious server can do to a software agent's
memory. The read-write state variables of an agent may be completely altered
by the mahcious server; thus, an agent that has just left the malicious server can
not trust any of its memory: All information collected prior to this point it time
— including data from servers visited prior to visiting the malicious server —
are suspect. Thus, only the results of computation done by those servers from
the {maximal) honest suffix of the agent's route, assuming that the computation
is independent of any input from previous servers, should be trusted.

Server Replication In [23], Minsky et. al. developed a general method for
mobile agent computation security, maxrying some ideas from the fields of fault
tolerance and cryptography. They propose that servers should be replicated,
and that replicated agents on these servers can use voting and secret shar-
ing/resplitting to move from one phase of the computation to the next.
Unfortunately, the fault model assumed in the paper is completely unreal-
istic: it assumes that replicated servers fail independently. In our Fly-By-Night
Airlines example, all replicated www.llybynight.comservers are under the ad-
ministrative control of flybynight.com, and malicious attempts to brainwash
software agents would occur on all of these servers. And while bribery of in-
dividual administrators of replicated servers by an outside adversary might
be independent events, bribery of the software engineers responsible for the
www. f l y b y n i g h t . com Web site is a much more likely scenario. Even if we assume
that Fly-By-Night Airlines is trustworthy, replicated servers in the real world are
likely to consist of identical hardware running copies of the same software: any
security holes found by an external attacker that allows him/her to compromise
one of the replicated servers is very likely to permit him/her to compromise all
the servers.

Agent Replication While the general approach proposed by Minsky et. al. fails
to be convincing, in certain special cases the fault tolerance style of approach can
solve or at least ameliorate the mobile agent security problem. Because server
replication does not help to reduce the risk of agent brainwashing, in the following
we will assume that there is only one server per administrative/security domain,
or when there are multiple servers in a domain, they are indistinguishable.
Consider the case where there is at most one malicious server in our airfare
minimization example. Assume that secure communication links exists between
the servers, and that the users possess individual certified public keys; servers
 265

may use these keys to verify the origin of the software agents. (Secure commu-
nication channels may be constructed cryptographically if servers also possess
cryptographic keys to authenticated and encrypted data among the parties as
needed.) Because we are assuming that there is only one dishonest server, we
know that the agent must stay on the circuit prescribed during agent configura-
tion.
Suppose we chose some sequence of servers 5 = si, S2,... , s„. We configure
two software agents Ai and A2, where ^1 will travel along S, and A2 will travel
over 5"^ = s „ , s „ _ i , . . . ,si.
Recall that we are assuming at most one malicious server. The all-honest
servers case is trivial, so we can ignore that case; henceforth we will assume that
there is exactly one bad server. Without loss of generality, assume that server
Sj is malicious, and that Sj is run by the airline with the lowest fare {j < i).
Furthermore, we assume that the malicious server will not attempt denial-of-
service attacks — it may do so by killing the software agent or by implanting
the belief that the lowest fare is offered by some third server which will later
repudiate this idea.
First, consider the j < i case. A\ will encounter the lowest-fare server (sj)
first, and when it arrives at Sj, its memory of the lowest-fare seen-so-fax may be
altered. When A\ returns with its result, it will report either Si as the server
with the lowest fare, or some Sk where fc > i if the malicious server did not
declare a fare lower than one that the agent will see later in its travels.
A2, on the other hand, will encounter the lowest-fare server after visiting the
malicious server. It will report the correct minimum price — since we assume
no denial-of-service attacks, the corrupt server will not have made this agent
believe that a (false) lower price exists elsewhere — and when A2 returns to the
user, the user will be able to determine the true minimum airfare.
Next, consider the j = i case. When this occurs, the malicious server can
alter its price to be just below that of the second lowest price offered and still
get business. This corresponds to a Vickery auction or second-price auction,^
except the situation is upside-down: instead of the highest bidder paying the
second highest bid price to obtain the goods being auctioned off, we have the
lowest airfare offer selling tickets at the second lowest quoted price. Note that
Vickery-style price determinations may be a desirable economic design choice
anyway, since Vickery auctions are designed to maximize the flow of pricing
information so bidders have no economic interest to hedge and not bid (and
reveal) the true prices that they are willing to pay.
The above agent-replication approach provided a partial solution for a special
case — at most one malicious server — the solution did not quite work "prop-

We do not have sealed bids here since the minimization is done by the agent; aji
alternative design would be to gather bids encrypted using the pubUc key of the
agent originator, preventing servers from knowing each other's prices directly. Of
course, servers could send out their own agents to discover such "commodity" prices;
this may have to be done through anonymizing proxies if the pricing could depend
on the consumer's identity.
 266

erly" to compute the true minimum airfare: when j = j , we could only achieve
second-best pricing, where what we obtain is the second-best airfare minus e.
Arguably, since airline servers may also send out agents to determine pricing at
other airlines — assuming price information can be obtained anonymously or
in such a way that we are assured that it is independent of consumer identity
(or race or age or ...) - Vickery pricing may be the end effect whenever there is
great consumer price sensitivity in any case. Applying some basic cryptographic
techniques, however, we can do a httle better.

3.4 Cryptographic Approaches

There are three cryptographic techniques that apply or may apply. The first uses
per-server digital signatures to vouchsafe partial results; the second is the use
of state authentication codes to improve on the fault-tolerance solution above;
and the third is the use of probabilistically checkable proofs (PCP) (a.k.a. "holo-
graphic proofs") or computationally sound proofs (CS proofs) to show that the
computation at the servers ran correctly. Note that these techniques deal with
the integrity of the result of computations at various servers — if the privacy of
the result is needed, the result can simply be encrypted using the originator's
public key.

Digital Signatures The application of digital signatures to mobile agents is

to use the signatures to vouchsafe partial results from computation done while
executing in a server. Going back to the airfare minimization example, what we
will do is to have each airline server sign a message of the form "this is the best
airfare found by your software agent at this server at this time" ? The message
signature is done using the airline's key, so it is non-repudiable and unforgeable.
The message may optionally be encrypted by the originator's public key to ensure
privacy.
The key observation is this: due to the unforgeability property, a malicious
server can not completely brainwash the agent — at worse, it can make the
agent forget the lowest airfare, which will be detected when an enumeration of
the signatures show that one airline's quote is missing.
By using digital signatures, we will either obtain the true minimum airfare
or be able to detect any tampering with the agent. The cost is that the agent
state grows linearly with the number of servers visited, where the fault-tolerance
approach required constant space (though the result was Vickery pricing rather
than true minimum). This is an acceptable overhead for most applications. Like
the fault-tolerance approach, however, the digital signature approach is not fully
general: it only applies only certain classes of functions where there are interme-
diate results that can be "compressed" (e.g., in this case, the intermediate result
is the best price found on the server — we don't care how much work was done
It is important to note that the query as well as the answer is signed and times-
tamped. Otherwise signed answers to the wrong question, asked by a different agent
dispatched by the adversary, could be substituted in place of the expected answer.
 267

in querying the airline's databases prior to finding this result, and we don't need
to prove that it took place and that it ran correctly).

Partial Result Authentication Codes The idea of a Partial Result Authen-

tication Code (PRAC) is very similar to that of a message authentication code
(MAC)[4, 5,17,19]. Instead of authenticating the origins of a message, we are
demonstrating the authenticity of an intermediate agent state or partial result
that resulted from running on a server. Similar to MACs, PRACs are cheaper
than digital signatures to compute, and have slightly different security proper-
ties.
The property that PRACs ensure is forward integrity: if a mobile agent vis-
its a sequence of servers S = Si, S j , . . . ,««, and the the first malicious server
is Sc, then none of partial results generated at servers Si, where i < c, can be
forged. This contrasts with a simple digital signature, where if an attacker com-
promises the generating host where the signature key is stored, the authenticity
of all messages signed with that key becomes questionable. The use of a digital
timestamping service [14] can have similar properties, except that in that case
a trusted third party (the timestamping service) is required and the granularity
of the timestamps limits the maximum rate of travel for the agents — the agent
must stay on a server until the next timestamping epoch before migrating to the
next server.
The notion of forward integrity is a generalization of the notion of "forward
secrecy", and have applications in other contexts [7]. The notion of forward
secure cryptography applies to pseudo-random number generation as well as
secrecy and integrity [6].

Simple MAC-based PRACs To have an agent use simple PRACs, we provide

the agent with a list of secret PRAC keys at agent dispatch, with a key per
server visited. Before leaving a server for the next, the agent summarizes its
partial results from its stay at this server in a "message" back to the agent
dispatcher. This message need not be sent back to the dispatcher immediately;
instead, it may be carried with the agent as part of its migrating state for later
transmission. This may be delayed so that it is "sent" to the dispatcher when the
agent returns. Alternatively, these messages may be "batched" and sent when the
networking bandwidth is cheap or available, e.g., when the dispatching mobile
host has reconnected to the network. To provide integrity, a MAC is computed
on this message, using the key associated with the current server; the message,
along with its MAC, comprises the PRAC. The critical difference between MACs
and PRACs is that after a PRAC is computed, the agent takes care to erase the
PRAC key associated with the current server prior to migrating to the next
server.
The ercisure step provides a very important security property: the partial
results from the honest prefix servers can not be modified. (This contrasts with
the honest suffix property from section 3.3, where the partial results from honest
servers visited after visiting the last malicious server are known to be unmod-
ified.) Suppose an agent A traverses a sequence of servers S = s i , S 2 , . . . ,s„,
 268

where at each server s, a partial result PRi is computed using key fcj, and
servers si,.. .Sj are honest and do not expose the internal state of the agent,
then 'ii,k : i < j < k,Sk can not forge PRj, since Sk. must know k, to change
PRi.

MAC-based PR AGs with One- Way Functions An obvious enhancement to simple

PRACs is to use a single key instead of n PRAC keys and use a m-bit to m-bit
one-way function to generate the list of PRAC keys. When the agent is initially
sent to server si, it contains key ki. When the agent prepares to go from server
Si to server Sj+i, it computes ki+i = f{ki), where / is a one-way function,
and erases all knowledge of fcj. As before, a server Sk can not forge PRACs
from previous servers, since it would have to break the one-way assumption to
determine the previous PRAC keys or break the MAC function.
More generally, instead of a m-to-m-bit function, an m to r bit one-way
function may be used. (Typically r < m.) In this case, to obtain ki+i from ki,
we simply use some (perhaps pseudo-random) known (m — r)-bit string t and
set fci+i = f{ki\t), where "|" denotes concatenation. If the probability that any
algorithm, given y, will find a pre-image x : f{x) = y is at most 2~"*-l-e, knowing
the last {m — r) bits of the pre-image x = ki\t should not help: if knowing t gives
an algorithm a probabihty of finding a pre-image of p > 2^^ + 2'"~''e, then
by guessing the value of t values (probability 2'""'"), we obtain an algorithm
which will find a pre-image with probabihty 2''"™p > 2"™ -I- e, contradicting our
one-way assumption.

Publicly Verifiable PRACs The MAC-based PRACs above required that the
agent originator maintained a secret key or keys in order to detect tampering
with the partial results. An obvious question is whether forward integrity can
be provided such that the integrity verification may be public — so that an
untrusted intermediate server not sharing the secret key with the originator
may nonetheless help detect tampering.
Like MAC-based PRACs, publicly verifiable PRACs are implemented by rely-
ing on the destruction of information when agents migrate. Here, we use a digital
signature system: when the agent is dispatched, it is given a list of secret sig-
nature functions sigi(m),... , sig„(m), along with usre-generated certificates for
their corresponding verification functions verifi(m, s ) , . . . , verif„(m, s). The ver-
ification functions would be signed by the user's signature function siguser('^).^
Like simple MAC-based PRACs, we use sigj(m) to sign the partial result com-
puted on server Sj, and erase sigj(m) prior to migrating to server st+i.
Similar to one-way function MAC-based PRACs, we can also defer key gen-
eration, so that most of it is done on the servers, which presumably have greater
resources. Here, the agent is given an initial secret signature function sigi(m)
and a certified verification predicate verifi(m, s); the signature function is used
both to sign partial results and to certify new verification functions.

* Signing a function verifi(.) simply consists of signing the parameters that specify the
function; in RSA, it would be the two values e,;,ni.
 269

The verification predicate verifj (m,s) (and its certificate) is public, and the
signature function sigi(m) is secret. When the agent is ready to leave server si,
it signs the partial result ri by computing sigi(ri). Next, it chooses (randomly)
a new signature / verification function pair from the signature system, sig2(m)
and verif2(m, s), and computes sigi (verifa) to certify the new signature functions.
Lastly, before the agent migrates to server S2, sigj is destroyed.
To use pubhcly verifiable PRACs, the list of certified verification predicates
must be either published and/or carried with the agent. When these predicates
are available with the agent, publicly verifiable PRACs enjoy an important prop-
erty not available with MAC-based PRACs: while at server Sj, the agent can
itself verify the partial results obtained while at servers Si, where i < j . In par-
ticular, this means that computations that depend on previous partial results
can detect any integrity violation of those results — the agent's computation can
abort early, instead of having to finish the computation and detecting integrity
violation only when the agent results return to the agent originator.

Proof Verification We would like to get a guarantee that the agent's com-
putation was done according to program specified in the agent. One possibility
is to forward the entire execution trace — or a signature of it [28] — to the
originator, who checks the trace if there is cause for suspicion. This however is
too costly. Furthermore, if only a signature is sent to commit to a trace, it is
unclear how to decide when to actually request the entire trace and then check
it. We would like to explore the use of holographic proof checking techniques [3].
This is quite a speculative idea. The current approaches are very theoreti-
cal. In principle they do help, but the cost in practice of existing solutions is
prohibitive. We are considering investigating ways to use the ideas in a more
practical way. Let us describe the ideas and issues to see what it is about.
Call the program x. Let y denote an execution trace. Define the predicate
p{x, y) to be 1 if this trace is correct (corresponding to running x) and 0 other-
wise. The server does not want to send y. But it can encode y as a, holographic
proof y'. This has the property that one needs to look at only a few bits of
y' to check that p{x,y) = 1. It is tempting from this to think that the server
can just transmit a few bits. But this does not work. The model necessary for
holographic proofs is that the verifier have available a fixed, "committed" proof
string y' that he can access at will. He will pick a few random positions here and
check something. So there is no choice but to transmit y' in entirety. We will not
save bandwidth. We will gain something: the verification process is faster. (The
verifier receiving y' will perform some quick spot-checks).
A better approach is to use computationally sound (CS) proofs as in [20,22].
Having constructed the holographic proof y' as above, the server hashes it down
via a tree hashing scheme using a collision-resistant hash function h. Only the
root of the tree is sent to the originator. This is relatively short, so bandwidth
is saved. In addition, certain challenges are implicitly specified by applying an
ideal hash function to this root, and the server also provides answers to them.
The total communication from server to originator is still small compared to the
 270

length of the original execution trace y, yet some confidence in the correctness
of y is transmitted!
The tree hashing is actually not impractical. What is prohibitive is con-
structing the holographic proof y' to which it is applied. This currently calls for
apphcation of NP-completeness techniques, including the use of the construc-
tion underlying Cook's theorem. What we might hope instead is to find a direct
holographic proof for the functions of interest, and then apply tree-hashing.

3.5 Trusted Environments (Secure Coprocessors)

An engineering solution to providing security for software agents is to build
a trusted / trustworthy execution environment for the agents. The Sanctuary
project is building such an environment to run within a secure coprocessor [18,
30,32], allowing Java-based agents to run securely; designing and implementing
the agent APIs needed to support mobile Java agents; and have developed the
basic technology — analysis of code using "local continuations" — by which
Java-based agents can migrate among unmodified Java interpreters running in
an secure-coprocessor environment [15].
In addition to the basic support for agent execution, the Sanctuary project
is developing the trust framework needed for inter-server communications. This
necessarily implies having some basic public key infrastructure — we will lever-
age off of the existing work being done to support SSL [12], PCT [8], and TLS
[11].^
A similar approach is found in [31]. Earlier work on analyzing the physical
security assumptions for using tamper-detecting hardware, the trust model(s)
involved, verification/inspection techniques, and the public key infrastructure
needed to support such a system is found in [32]. The analysis there used a
traditional operating system / application view instead of a more agent-centric
view; in our opinion, an agent system is simply another application, a set of
user-level servers which runs on top of a traditional security kernel within the
secure coprocessor.

4 Trust IVEodels
The issue of trust models is very important to agent-based computing. Agents do
not just need a trusted computing base (TCB) — trust may not be so binary in
nature. Instead, agents (or their deployers) may decide that it is okay to run in a
software-only environment if such an environment is hosted by a well-known and
trusted entity, but the use of physical protection to maintain the trustworthiness
of a trusted third-party provided execution environment is needed when the
environment is hosted by an entity with no reputation to protect and/or where
no legal remedies may be obtained.
* The Internet Engineering Task Force's Transport Level Security group have devel-
oped a merged protocol based on SSL version 3 and features from PCT. These
protocols require a merchant-side public key infrastructure.
 271

In Sanctuary, we envision that the trust decision will be made by the agent's
software itself. Thus, trust specification is simply an object in Java, and any
effectively computable function may be used. This is similar in spirit with the
work of Blaze and Feigenbaum [10], except that by unifying the agent language
and the trust specification language, the programmer's work is simplified.

5 Mobile Java

Other approaches to providing mobility to Java programs [26] requires modifying

the interpreter. In Sanctuary, we intend to provide a mechanism to migrate Java-
based agents that can run on unmodified Java interpreters [15]. This strategy
will enable wider acceptance of mobile agents, leveraging off of the work done by
Sun/Javasoft and other companies. This means that just-in-time compilation as
well as compilation of bytecode to native code are viable alternatives to bytecode
interpretation for our system.

6 Future Work

The Sanctuary project group are examining important security issues in mo-
bile agent computing. This paper has discussed some preliminary results and
directions.
The primary goal is to build a secure agent environment insofar as it is the-
oretically feasible, including issues such as providing clean abstractions to make
programming errors less likely as well as good cryptographic support. Currently,
we are building a trusted Java agent environment to run within a secure copro-
cessor and designing APIs that permit agents to exist both in a hardware-based
secure environment and in a software-only environment unchanged (but permit-
ting security property queries). Next, we will build the necessary software tools to
permit Java-based agents to be mobile. Our techniques, once implemented, will
enable these agents to run on unmodified Java interpreters; this design approach
permits greater acceptance of our work, since no complex installation process
will be required, and it will allow our system to track new Java releases more
easily. Additionally, we are examining alternative methods for providing security
for software agents through fault tolerance and cryptographic approaches (e.g.,
distributed function evaluation, additional uses of digital signature techniques,
etc).

Acknowledgement s
The author wishes to thank Mihir Bellare for his invaluable help in preparing
this paper.
This research was funded in part by a National Science Foundation CAREER
Award (CCR-9734243), a Faculty Development Award from National Semicon-
ductor Corporation, and a gift from the Powell Foundataion.
 272

References
1. Martin Abadi and Joan Feigenbaum. Secure circuit evaluation. Journal of Cryp-
tography, 2(1):1-12, 1990.
2. Martin Abadi, Joan Feigenbaum, and Joe Kilian. On hiding information from an
oracle. Journal of Computer and System Science, 39(l):21-50, August 1989.
3. Laszlo Babai, Lance Fortnow, Leonid A. Levin, and Mario Szegedy. Checking
computations in polylogarithmic time. In Proceedings of the Twenty Third Annual
ACM Symposium on Theory of Computing, pages 21-31, New Orleans, Louisiana,
May 1991.
4. Mihir Bellare, Ran Canetti, and Hugo Krawczyk. Keying hash functions for mes-
sage authentication. In Neil Koblitz, editor, Advances in Cryptology — Crypto '96,
volume 1109 of Lecture Notes in Computer Science. Springer-Verlag, 1996.
5. Mihir Bellare, Roch Guerin, and Phillip Rogaway. XOR MACs: New methods
for message authentication using finite pseudo-random functions. In Advances in
Cryptology — Crypto '95, volume 963 of Lecture Notes in Computer Science, pages
15-28. Springer-Verlag, 1995.
6. Mihir Bellaxe and Bennet Yee. Forward-secure cryptography: How to protect
against key exposure. Work in progress.
7. Mihir Bellare and Bennet Yee. Forward integrity for secure audit logs. Technical
report, Computer Science and Engineering Department, University of California
at San Diego, November 1997.
8. Josh Benaloh, Butler Lampson, Terence Spies, Dan Simon, and Bennet S. Yee.
The PCT protocol, October 1995.
9. Brian N. Bershad, Stefan Savage, Przemyslaw Pardyak, Emin Giin Sirer, Marc E.
Fiuczynski, David Becker, Craig Chambers, and Susan Eggers. Extensibility, safety
and performance in the spin operating system. In Proceedings of the Fifteenth
Symposium on Operating Systems Principles, December 1995.
10. Matt Blaze, Joan Feigenbaum, and Jack Lacy. Decentralized trust management.
In Proceedings 1996 IEEE Symposium on Security and Privacy, May 1996.
11. Tim Dierks and Christopher Allen. The TLS protocol, version 1.0,
November 1998. Internet Engineering Task Force Internet Draft; see
http://www.ietf.cnri.reston.va.us/internet-drafts/draift-ietf-tls-protocol-06.txt.
12. Alan Preier, Philip Karlton, and Paul Kocher. The SSL protocol version 3, De-
cember 1995.
13. J. Steven Pritzinger and Marianne Mueller. Java security, 1996. Published as
http://www.javasoft.com/security/whitepaper.ps.
14. Stuart Haber and W. Scott Stornetta. How to time-stamp a digital document.
Journal of Cryptology, 3(2), 1991.
15. Matthew Hohlfeld and Bennet S. Yee. How to migrate agents. Techniccd Report
CS98-588, Computer Science and Engineering Department, University of California
at San Diego, La Jolla, CA, August 1998.
16. Wilson C. Hsieh, Marc E. Fiuczynski, Charles Garrett, Stefan Savage, David
Becker, and Brian N. Bershad. Language support for extensible operating sys-
tems. In Proceedings of the Workshop on Compiler Support for System Software,
February 1996.
17. IBM Corporation. Common Cryptographic Architecture: Cryptographic Application
Programming Interface Reference, SC40-1675-1 edition.
18. IBM 4758 PCI cryptographic coprocessor. Press Release, August 1997.
http://www.ibm.com/Security/cryptocards/.
 273

19. R. R. Jueneman, S. M. Matyas, and C. H. Meyer. Message authentication codes.

IEEE Communications Magazine, 23(9):29-40, September 1985.
20. Joe Kilian. A note on efficient zero-knowledge proofs and arguments. In Proceedings
of the Twenty Fourth Annual ACM Symposium on Theory of Computing, Victoria,
British Columbia, Canada, May 1992.
21. Steven McCanne and Van Jacobson. The BSD packet filter: A new architecture
for user-level packet capture. In USENIX Technical Conference Proceedings, pages
259-269, San Diego, CA, 1993. USENIX.
22. Silvio Micali. CS proofs. In Proceedings of the 35th IEEE Symposium on Foun-
dations of Computer Science, pages 436-453, Santa Fe, New Mexico, November
1994.
23. Yaron Minsky, Robbert van Renesse, Fred B. Schneider, and Scott D. Stoller.
Cryptographic support for fault-tolerant distributed computing. Technical Report
TR96-1600, Department of Computer Science, Cornell University, July 1996.
24. George Necula. Proof carrying code. In Proceedings of the Twenty Fourth Annual
Symposium on Principles of Programming Languages, 1997.
25. George Necula and Peter Lee. Safe kernel extensions without run-time checks.
In Proceedings of the Second Symposium, on Operating Systems Design and Im,ple-
mentation, Seattle, WA, October 1996.
26. Mudumbai Ranganathan, Anurag Acharya, Shamik D. Sharma, and Joel Saltz.
Network-aware mobile programs. In Proceedings of the Usenix 1997 Annual Tech-
nical Conference. Usenix, 1997.
27. Tomas Sander and Christian F. Tschudin. Protecting mobile agents against mali-
cious hosts. In Giovanni Vigna, editor, Mobile Agents and Security, volume 1419 of
Lecture Notes in Computer Science, Start-of-the-Art Survey, pages 44-60. Springer-
Verlag, 1998.
28. Giovanni Vigna. Cryptographic traces for mobile agents. In Giovanni Vigna, editor,
Mobile Agents and Security, volume 1419 of Lecture Notes in Computer Science,
State-of-the-Art Survey, pages 137-153. Springer-Verlag, 1998.
29. Robert Wahbe, Steve Lucco, T. E. Anderson, and Susan L. Graham. Efficient
software-based fault isolation. In Proceedings of the ACM SIGCOMM 96 Sympo-
sium. ACM, 1996.
30. Steve R. White, Steve H. Weingart, William C. Arnold, and Elaine R. Palmer.
Introduction to the Citadel architecture: Security in physically exposed environ-
ments. Technical Report RC16672, Distributed security systems group, IBM
Thomas J. Watson Research Center, March 1991. Version 1.3.
31. U. Wilhelm, S. Staamann, and L. Buttyan. Introducing trusted third parties to
the mobile agent paradigm. In Secure Internet Programming, Lecture Notes in
Computer Science, Springer-Verlag Inc., New York, NY, USA, 1999.
32. Bennet S. Yee. Using Secure Coprocessors. PhD thesis, Carnegie Mellon University,
1994.
 Mutual Protection of Co-operating Agents

Volker Roth

A b s t r a c t . Security is a fundamental precondition for the acceptance

of mobile agent systems. In this paper we discuss protocols to improve
agent security by distributing critical data and operations on mutually
supporting agents which migrate in disjunct host domains. In order to
attack agents, hosts must collude/conspire across domains. Proper selec-
tion of itineraries can minimize the risk of such coalitions being formed.

K e y w o r d s : mobile agent security, malicious host, distributed

applications, itinerciry recording, agent cooperation, ecash

1 Introduction

Mobile agents are bundles of program and state, that move within a
network to perform tasks on behalf of their owners. The benefits offered
by mobile agents unfold in areas where it is advantageous to move the
computation process over a network to the source of data instead vice
versa; for instance if huge amounts of data must be processed, or the
network is slow, expensive, or is not permanently available. Among the
manifold uses for mobile agents, electronic commerce applications are
noted most frequently. With good reason - electronic commerce provides
perfect grounds to illustrate the benefits of mobile agents as well as the
threats that keep us from using them in open networks, so far.
One example benefit of a mobile agent is its ability to carry out certain
tedious or time-consuming tasks autonomously while its owner is offline.
The goal of, for instance, a shopping agent might be finding the best
offer for a given product description, thus optimising the benefit to its
owner (which is legitimate). However, shops might be tempted to opti-
mise for their own benefit, even if it means optimising at the expense of
the agent in unfair ways (which is not legitimate). This includes manipu-
lation of offers previously collected by the agent, as well as abusing agents
as mediators of attacks on competitors. Hence, neither party trusts the
other, and both share their distrust in the interconnecting network. As a
 276

consequence, the security requirements for mobile agent systems are man-
ifold as well as demanding. Quoting from [2]: "/it is difficult to exaggerate
the value and importance of security in an itinerant agent environment.
While the availability of strong security features would not make itiner-
ant agents immediately appealing, the absence of security would certainly
make itinerant agents very unattractive."
In general, the problem of malicious hosts is considered particularly chal-
lenging. The threats imposed by malicious hosts are immanent in the
way mobile agent systems are built. Since agents are executed on their
host, each instruction in the agent's code is observed by the controlling
(virtual) machine which also maintaines the agent's state. This causes a
number of security concerns of which the most prominent ones are:

1. The integrity of the agent, in particular the integrity of its mutable

part, need to be protected. Agents might be abused as innocent car-
riers of illegal or offensive materials such as warez (pirated software).
Hosts may also try to delete, replace, or invalidate commitments to the
agent, such as terms negotiated in electronic commerce applications.
This would enable the host repudiate said terms later.
2. Maintaining the secrecy of the agent's computations and data is a
fundamental requirement for fair negotiations as well as for computa-
tions on confidential information such as the preferences (or profile)
of the agent's owner, or secret keys.
3. Protecting the integrity of the agent's control flow is a precondition for
any agent to trust its own decisions. Otherwise, malicious hosts might
make agents believe that an offer is acceptable when it is actually not.

Notable advances were made with regard to items 1 and 2. Karjoth

et al. [3] introduced the notion of strong forward integrity and proposed
protocols for protecting the computation results of free-roaming agents.
Their work is an extension of partial result authentication codes intro-
duced by Yee [15]. Roth and Jalali proposed an agent structure that
supports access control and authentication of mobile agents [5]. Agent au-
thentication and state appraisal is covered by Berkovits et al. [8]. Sander
and Tschudin [6] introduced the notion of mobile cryptography and de-
vised approaches for non-interactive computing with encrypted functions
using homomorphic encrytion schemes. Tschudin also proposes an ap-
proach for securing termination signals for mobile services [11] which
incorporates ideas from Riordan and Schneier [4]. Regarding item 3, Vi-
gna [12] proposed cryptographic traces to create verifiable execution traces
 277

of agents. Vigna also discusses a number of drawbacks with his approach

such as hmitation to single-threaded agents, extensive computational and
memory overhead, as well as limitation to a posteriori detection. The no-
tion of trust, its role in mobile code systems, as well as trust management
is discussed for instance by Swarup et al. [10] and Blaze et al. [1].
The majority of risks stem from the fact that agents are deployed in
open untrusted networks. However, even the hosts connected through
this network do not know and do not necessarily trust each other. In the
remainder of this paper we intend to exploit exactly this fact for improving
the security of mobile agents against attacks by malicious hosts. The
general idea is based on the co-operation of multiple agents with the
goal of mutual protection (of the application built on them). Thus, co-
operation of agents is used to establish a distributed "virtual shelter"
inside the open network, in which agents can hide and to which they may
retreat while assuring the security of their counterparts.
In Section 2 we motivate and formalise this idea, and give definitions and
notations used in our paper. In Section 3 and Section 4 we devise two pro-
tocols, based on co-operating agents, suitable to protect different aspects
of the general purchase agent against malicious hosts. One protocol is
directed at recording the actual itinarary taken by a free-roaming agent;
the second protocol addresses the delegation of electronic payments to
mobile agents. Conclusions will be drawn in Section 5.

2 Co—operating Agents

With regard to mobile agent security, the most conservative approach

is to assume that each host on a mobile agent's itinerary is hostile and
willing to collaborate with other malicious hosts visited by the agent on
its route. This assumption is as realistic as the assumption that hosts can
be generally trusted, though. A more reasonable assumption is probably
the following:

Given a particular mobile agent, at any point in time, a certain

percentage of hosts might be malicious.
Not all malicious hosts are willing to collaborate with other hosts
in attacking a mobile agent.

The percentage of malicious hosts likely depends on the gain which can
be expected from successfully attacking mobile agents weighted against
 278

the costs of mounting the attack as well as the risk of detection and the
consequences of being detected. Collaborations of multiple hosts on an
agent's itinerary yields more power but it also requires close coordina-
tion and increases the danger of leaks which might lead to disclosure of
the collaboration. Whether an agent will be attacked by a single mali-
cious host or a collaboration of hosts on its itinerary depends on a sea
of unpredictable parameters that are unique for each instantiation of an
agent.
Still, we would like to model partitions of hosts based on their willingness
to collaborate with regard to attacking a fixed agent. For this reason, let
'H be the set of hosts interconnected by a network. For a given instan-
tiation of an agent let 7?. be a relation defined as TZ C y. x y. with the
interpretation (h,:,hj) & TZ <^ hi and hj collaborate in attacking the agent.
Let Ha,Hb be non-empty subsets of Ti with (HQ X H;,) n 7?. = 0. These
two sets are denoted non-colluding. A special host is the first host (the
origin) of an agent, since this needs to be a trusted host.
Two co-operating agents are defined to be agents a and b such that the
itinerary of a includes only hosts in H^ and ft's itinerary only includes
hosts in Hf,. Let ha and h^ be the computing environments currently
executing agents a and b respectively. Occasionally, we will say that ha
is "the host of" agent a, or "ha is agent a's host".
Although agent a might be attacked by host ha, by definition this host
may not attack the co-operating agent b without breaking into host h^.
This can be exploited to design protocols for securing agents against at-
tacks by single malicious hosts as well as hosts that collaborate with other
hosts as long as the collaboration does not span the itineraries of both
agents.
A number of strategies can be used as starting points for developing said
protocols:

Secret sharing: Authorisation data is secretly shared by both agents.

A single share conveys no information about the shared secret. There-
fore, a malicious host must get the remaining share by asking a host
on the co-operating agent's itinerary to steal it, or it must break into
that host.
R e m o t e authorisation: The decision whether the share of the autho-
risation data is passed to the host executing the agent is taken by the
co-operating agent. The agent itself prefilters the data upon which the
 279

decision is based, and transfers it to its co-operating agent. Therefore,

a malicious host must ask the remote host to manipulate the decision,
or it must break into that host.
Remote storage of cominitments: An agent transfers commitments
of the host on which it runs, such as commercial offers, to its co-
operating agent (probably in the course of remote authorisation),
which verifies and stores it for future reference and non-repudiation.
In order to undo or invalidate its commitment, a malicious hosts needs
to break into a host on the co-operating agent's remaining itinerary.

In the remaining sections we illustrate the idea of mutual protection of

co-operating agents by giving example protocols. For those protocols to
work we need to make two additional assumptions:

- Hosts transport agents through authenticated channels.

- Hosts provide an authenticated communication channel to the two
cooperating agents.
- The authenticated identity of the remote host, the authenticated iden-
tity of the host the agent came from as well as the local host's identity
are provided to the hosted agent.

Yee already pointed out that "if an agent is running on an honest server,
both these answers (for the peer identity and the local host's identity)
will be correct..." [15]. We assume that a host is honest unless it may
successfully attack an agent on its own, or with the help of other hosts
on this agent's itinerary. In other words, we assume that hosts do not
randomly introduce lies.

3 Tracing Loose Routes

A simple yet effective attack on a mobile agent is to not let the agent
migrate to the servers of competitors. This particularly affects mobile
agents with loose itineraries in comparison to agents whose itineraries
are defined a-priori, because deviations from a fixed itinerary are easier
to spot and prove.

We would like to record the actual loose route taken by a free-roaming

agent without any possibility of manipulation by the hosts on its route.
 280

Let a and b be two co-operating agents, and let B.afibQ'H be two non-
colluding sets of hosts. Both agents shall return to their origin upon com-
pletition of their tasks. Each agent b records and verifies the route of its
co-operating agent a as described below.

Definition: Let h^eHa be the i host being currently visited by agent a

and let id(hi) be the identity of host hj. Let prevj be agent a's idea
of the identity id(hj_i) of its previous hop. nextj shall denote the
identity of the next hop agent a wants to take while being on host h^.
The agents start at host ho, hence h„=ho for a route with n hops.
Initialization: Let ho be the origin of agents a and b. ho has to be a
trusted host with respect to a and b. For agents a and b, nexto is
set to the first hop of their respective itineraries. Both agents are
subsequently sent to their first hop.
S t e p i,2 G { 1 , . . . , n } : Agent a sends a message containing the next
hop nextj and the previous hop prev^ to agent b. The authenti-
cated channel enables Agent b to learn id(hi). Agent b verifies that
id(hi)=nexti_i Aprevi=id(hi_i) and appends next^ to the stored route.

Security of the protocol: It is straightforward to see that if host hj

forwards agent a to a host hj+i with id(hi4-i)7^nexti then host hj+i must
either successfully masquerade as the host with id next^ or it has to deny
communication between the co-operating agents. On the other hand, if
host hj+i permits the communication and properly authenticates itself (in
other words, h^+i is honest regarding the protocol) then agent b discovers
that host hi sent agent a to the wrong destination.
If id(hi+i)/nexti then host hj+i cannot put agent o back on its route by
sending a to the host with identity next^ because agent b recorded id(hi)
as prevj+i. As a matter of fact, hj+i must either be honest (identifying hj
as a cheater in the process) or hj+i has to collaborate with h^ in putting
the agent back on its expected route (more precisely, h^ has to put back
the agent on its route itself, or it must disclose its authentication keys
to hi^_i). The last case reduces to either simply sharing a copy of the
agent with hj+i, or merging hosts hj and hj+i into a single host under
the identity of hj - hardly something which can be prevented or detected
at all.
A malicious host h^+i might incriminate a honest host h, by claiming to
have received agent a from some other host h', hence implicating that hj
sent a to h' instead of the host with identity nexti=id(hi+i). The protocol
 281

is not able to decide which one of the two hosts is the culprit. However, if
hi+i really received agent a from h' then hj+i should be able to produce
a copy of a which is signed by h' given some additional agent protection
mechanisms are implemented (see [5]).
If agent a is killed, one of two hosts might be responsible and the protocol
cannot decide which one. In addition to that, some host hi-).i might take
two agents ai and a2 both being received by the same host hj and switch
the recording of the route of ai to agent 62 and vice versa. Therefore,
co-operating agents should also exchange and verify (unique) identity
information that is bound to the agent's static part by their owner's
signature (see [5]). In that case, attempts to send fake ids are detected
on the first honest host. The protocol must be enhanced accordingly.

4 Electronic Cash Payments

Current descriptions of mobile purchase agents primarily address collect-

ing and filtering of offers. In this section, we would like to consider dele-
gating payment authorisation to mobile agents as well. The risks in doing
so are obvious. Digital representations of money may be copied, payment
decision altered, and commitments erased after payment, by malicious
hosts. However, modelling a complete purchase cycle with mobile agents,
including payment authorisation, is an attractive thought. Cash-like sys-
tems are particularly well-suited because in contrast to for instance credit
card based systems the possible loss is limited to the value of the digital
coin.
Co-operating agents provide an approach to tackle security problems in-
volved in payment authorisation. Protocols such as Chaum's digital elec-
tronic cash protocol (as described for instance in [7]) can be adapted
to work with co-operating agents as described in the protocol given be-
low. Chaum's protocol provides detection of double spending with cheater
identification.
Let a and h be two co-operating agents, and let Ha,H(,C^ be two non-
colluding sets of hosts.

Initialisation The owner of the agents prepares a money order m as de-

scribed in [7]. The identity strings are prepared by randomly choosing
keys A;', A;[, ? = 1 , . . . , n and computing
{h = {EJ^l{IlH{I\)),E^[Il,H{Il)))\ i = l...n]
 282

where E is a, suitable encryption function and H is a, strong crypto-

graphic one-way hash function. The money order as well as the keys
are secretly shared between agents a and b. Both agents are sent on
their respective itineraries.
Step 1: Agent a locates a product or service on host ha (the one on which
the agent runs) that it wishes to purchase, and requests a signed offer
describing the subject and terms of the purchase.
Step 2: Host ha provides the requested offer as well as the selector string
defined in the ecash protocol.
Step 3: Agent a forwards the offer and the selector string to its co-
operating agent b.
Step 3: Agent b validates the offer's signature. In order to do so, it veri-
fies that the identity of the offer's signer matches the one determined
from the authenticated channel, and that the offer's signature is valid.
It goes on to verify the terms of the offer. If the verification fails then
a is notified of the result and the protocol is aborted.
Step 4: Agent b stores the offer (if ha later on repudiates having made
that offer then its signature may be used as a proof), transfers its
share of m to a, and opens the selected halves of the identity strings
by sending the appropriate key shares as well.
Step 5: Agent a passes the data on to ha.
Step 6: Host ha reconstructs the money order m, verifies the bank's
signature and makes sure that agent b properly opened the selected
halves of the identity strings. The ecash protocol goes on as described
in [7]. If the verification fails then the agents are notified, and the
protocol is aborted.
Step 7: Host ha delivers the purchased goods.

Security of the protocol: Clearly, host ha cannot manipulate the pay-

ment decision in step 3 without collaborating with h^. Neither of the
hosts involved can steal the money order without the help of the other
(or breaking into the remote host). Neither can host ha alter, invalidate or
delete its offer after step 4. Of course, either the host or the co-operating
agents may terminate the protocol at will. However, this is a general
problem of electronic payment protocols.
However, any two hosts may profit from a joint attack on the cooperat-
ing agents by sharing the additional wealth gained from defrauding the
agents compared to the profit gained from honest behavior. Therefore
the domains of both agents' itineraries must be chosen with great care
in order to assure (HQ, X H;,) n 7?. = 0 holds with reasonable confidence.
 283

Security can be improved by additional measures [5] which ensure that

payment can only be made by b while being on particular hosts.

5 Discussion and Conclusions

In this paper, we introduced the concept of co-operating agents as well

as the basic principles on which protocols for mutual protection of such
agents rest. These principles were illustrated by giving two example proto-
cols, one for recording the actual itinerary taken by free-roaming agents,
and one for protecting ecash payments with co-operating agents.
The ability to determine the actual route taken by an agent in principle
allows further improvements of the security of agents against unautho-
rised and unnoticed manipulation. Hosts may commit to changes in the
structure of an executed agent by digitally signing incremental logs of
the agent's state upon its migration. According to the recorded route,
the chain of signatures can be validated. Hence, in addition to strong
forward integrity [3] co-operating agents may provide resistance against
truncation and erasure with non-repudiation. Hosts which do not want
to forward agents to the servers of certain competitors might still do so.
However, the itinerary of the cooperating agent will always show this fact.
However, the ability of co-operating agents to mutually protect them-
selves relies crucially on the assumption that no pair of hosts chosen from
both agent's itineraries collude. This precondition must be ensured with
reasonable confidence. The most simple solution is to have one agent mi-
grate to a trusted host while the other agent is free to roam the network.
In this degenerate case, co-operating agents reduce to the trusted third
party concept with the important difference, that even the trusted third
party need not be fully trusted. Therefore, given a sufficient number of
agent servers in the Internet, even a randomised strategy might be suffi-
cient: One itinerary is randomly chosen and fixed while the second agent
is free to choose its next hop at run time. The randomised strategy is
based on the assumption that the sets of collaborating hosts are small
compared to the number of available hosts. The overall security might be
improved by complementing the approach taken in this paper with re-
search in agent routing policies such as the one indicated by Swarup [9].
In particular, the ecash adaption protocol illustrates how distributed ap-
plications based on mobile agents may be securely deployed while preserv-
ing in particular the autonomy of mobile agents as well as their ability to
 284

perform useful tasks off-line, and maintain confidence in the delegation

of authority as well.
In summary, protocols based on co-operating agents have merit since they
are less susceptible to attacks by coalitions of hosts than single agents.
The underlying principle is a generalization of the trusted third party
concept which is less restrictive and easier to meet.
We plan to implement and evaluate the protocols described in this paper
within the Java-based SeMoA platform (Secure Mobile Agents). SeMoA
denotes an ongoing development project undertaken by the Fraunhofer
Institut fur Graphische Datenverarheitung, with the goal of exploring and
developing techniques and architectures allowing secure deployment of
mobile agent technology in the area of multimedia and e-commerce ap-
plications.

6 Acknowledgements

I would like to thank the reviewers for their extensive comments on the
initial manuscript, which were most valuable for improving the quality of
this paper.

References
1. BLAZE, M . , FEIGENBAUM, J., IOANNIDIS, J., AND KEROMYTIS, A. he role of trust
management in distributed systems security. In Secure Internet Programming [14].
2. C H E S S , D . , GROSOF, B . , HARRISON, C , LEVINE, D . , PARRIS, C , AND TSUDIK, G .
Itinerant agents for mobile computing. IEEE Personal Comm,unications (October
1995), 34-49.
3. K A R J O T H , G . , ASOKAN, N . , AND GULCU, C . Protecting the computation results
of free-roaming agents. In Mobile Agents (MA '98), vol. 1477 of Lecture Notes in
Computer Science. Springer Verlag, Berhn Heidelberg, September 1998, pp. 1-14.
4. RIORDAN, J., AND SCHNEIER, B . Environmental key generation towards clueless
agents. In Mobile Agents and Security [13], pp. 15-24.
5. R O T H , V., AND JALALI, M . Access control and key management for mobile agents.
Computers & Graphics 22, 3 (1998). Special issue Data Security in Image Com-
munication and Networks.
6. SANDER, T . , AND TSCHUDIN, C . F . Protecting mobile agents against malicious
hosts. In Mobile Agents and Security [13], pp. 44-60.
7. SCHNEIER, B . Applied Cryptography, 1 ed. John Wiley &i Sons, Inc., 1994, section
6.7, pp. 120-122. Digital Cash Protocol # 4 .
8. SHIMSHON BERKOVITS, JOSHUA D . G U T T M A N , V. S. Authentication for mobile
agents. In Mobile Agents and Security [13], pp. 114-136.
 285

9. SwARUP, V. Trust Appraisal and Secure Routing of Mobile Agents. DARPA

Workshop on Foundations for Secure Mobile Code, Monterey, CA, USA, March
1997. Position Paper.
10. SwARUP, v . , AND FABREGA, J. T. Understanding trust. In Secure Internet Pro-
gramming [14].
11. TsCHUDiN, C. Apoptosis — the programmed death of distributed services. In
Secure Internet Programming [14].
12. ViGNA, G. Cryptographic traces for mobile agents. In Mobile Agents and Security
[13], pp. 137-153.
13. ViGNA, G. Mobile Agents and Security, vol, 1419 of Lecture Notes in Computer
Science. Springer Verlag, Berhn Heidelberg, 1998.
14. VlTEK, J., AND JENSEN, C . Secure Internet Programming: Security Issues for
Mobile and Distributed Objects. Lecture Notes in Computer Science. Springer-
Verlag Inc., New York, NY, USA, 1999.
15. Y E E , B . S . A sanctuary for mobile agents. In Secure Internet Programming [14].
 Part III

Implementations
 Access Control in Configurable Systems

Trent Jaeger

Abstract. In a configurable system, operating systems and applications

are composed dynamically from executable modules. Since dyneimically
downloaded modules may not be entirely trusted, the system must be
able to restrict their access rights. Current systems assign permissions
to modules based on their executor, provider, and/or name. Since such
modules may serve specific purposes in programs (i.e., services or appli-
cations), it should be possible to restrict their access rights based on the
program for which they are used and the current state of that program. In
this paper, we examine the access control infrastructure required to sup-
port the composition of systems and applications from modules. Access
control infrastructure consists primarily of two functions: access control
policy specification and enforcement of that policy. We survey represen-
tations for access control policy specification and mechanisms for access
control policy enforcement to show the flexibility they provide and their
limits. We then show how the Lava Security Architecture is designed to
support flexible policy specification and enforcement.

1 Introduction
A hope afforded by the mass connectivity of the Internet is that it would enable
users to engage in a variety of new applications regardless of their geographical
distance. Thusfar, the most successful applications to be enabled by the Inter-
net are client-server appUcations supported by the World-Wide Web. In these
applications, clients simply request information from servers. Servers may use
complex databases or dynamically determine the interface for processing the
request, but almost all the computing is done at the server.
While a number of useful applications can be developed using the client-
server paradigm, other computing approaches may be more appropriate for some
applications. This is particularly true for applications where the movement of
the computation is preferred over the movement of the data. In collaborative
applications, often the computation is smaller than the application state, so
each collaborator stores the application state and communicates their actions to
the group [16,26]. In this peer-to-peer approach, each collaborator may execute
an action performed by any member of the group on their own copy of the
shared state. Collaborative applications must be able to restrict the permissions
of such actions based on the state of the application. For example, only the files
currently being shared should be accessible to participants using a shared editor.
Second, research is underway to investigate how the modularity and exten-
sibility of operating systems can be improved. A number of systems have been
devised that can extend the operating system dynamically using kernel exten-
sions [4,12,42]. Also, the OSKit project is investigating how operating systems
 290

can be composed from system service components [14]. This module-on-demand

approach permits executable modules to be dynamically downloaded to sup-
port existing services or to compose services dynamically. The same argument
also can apply for the composition of applications. The system must be able to
restrict the permissions of such modules. This restriction may depend on the
purpose for which they are used. For example, a memory system extension for
an application should only obtain access to that application's memory objects.
Lastly, emerging applications, such as mobile agents, depend on the move-
ment of the computation from clients to servers [46]. This server-client approach
to computing permits a remote principal to act on server resources. In a com-
monly envisioned application, agents can perform complex queries on server
data to gather information for decision-making. Thus, agents must be able to
use server computing resources to read server data. Also, it may be permissible
in some instances for such agents to modify server state. Servers must be able
to restrict the permissions of its agents, particularly to system resources, but it
must also be possible to enable applications to control the rights available to the
agents they use.
The major impediment to the use of these computing approaches is an ad-
equate framework to support the enforcement of system access control policies
on dynamically loaded modules. At present, effective access control infrastruc-
ture for building such applications does not exist, so few applications use these
computing approaches, regardless of their advantages. We consider that access
control consists primarily of two functions: (1) access control policy specifica-
tion and (2) enforcement of those access control policies. Access control models
that support the specification of access control policies flexible enough to limit
modules based on their purpose are not used in practice. The typical approach
is to base permissions solely on the module author and name. Mechanisms for
enforcement of limited rights for modules has been investigated primarily in
language-specific systems [11,18,32,35]. The effectiveness of these mechanisms
are limited because they are depend on a large, complex trusted computing base
(i.e., the virtual machine, byte code verifier, etc.), are language-specific, depend
on unproven language separation features, do not control resource utilization,
and focus solely on protection of the system from the components. The access
control enforcement mechanisms need to be built into the operating system at
the lowest levels to be generally effective, but they must not significantly degrade
operating system performance.
In this paper, we survey representations for access control pohcy and mech-
anisms for access control policy enforcement that may be useful in building
configurable systems. In Section 2, we outline the problems in enforcing access
control on dynamically loaded modules. In Section 3, we examine the flexibility
and hmitations of different access control modeling features. The goal of this
discussion is to identify the access control representation features necessary to
express the access control policy for systems that use modules to compose appli-
cations and services. In Section 4, we examine the requirements of mechanisms
that can enforce flexible access control policies. The goal of this discussion is to
 291

derive a enforcement mechanisms that can support flexible and efficient enforce-
ment of access control policy. In Section 5, we describe the representations and
mechanisms chosen for the Lava Security Architecture.

2 Security Requirements
We assume a computing model in which principals (e.g., users and services)
consume resources (e.g., CPU and memory) to perform operations (e.g., read
and write) on objects (e.g., files and channels). The abihty to use resources to
perform operations on objects are called the access rights of a principal. The
fundamental problem posed by the computing approaches described above is to
restrict the access rights of each dynamically downloaded module consistently
with its service or application requirements. Since these requirements may vary
with each use, the job of access control policy specification and management
is daunting. Also, authorization mechanisms must be effectively and efficiently
able to enforce these requirements. In this section, we specify the access control
policy and authorization mechanism requirements for configurable systems.

2.1 Execution Model and Vulnerabilities

The execution model we envision for configurable systems is shown in Figure 1.

Remote servers store executable modules that are downloaded as necessary to
compose system services and applications. The modules may be written in any
language and/or compiled into the binary language of that machine. A module
may be loaded into an existing process with other modules or into a new process.
The system's trust in individual modules may vary widely. Some may be trusted
device drivers, but others may be untrusted to varying degrees. The system's
security policy is enforced by its trusted computing base (TCB). The TCB must
determine whether the module should be loaded, the process in which it is loaded,
the permissions of that process, and the ability of that process to obtain other
permissions.
Unfortunately, an untrusted binary module, unless controlled, can perform a
wide variety of attacks. Below, we list a taxonomy of common attacks:

— Unauthorized read: Read secret host data

— Unauthorized modification: Change host data or modify host binaries
(i.e., act as a virus)
— Unauthorized execution: Execute another program that gives the module
greater privileges (e.g., starting a shell)
— Unauthorized transmittal: Forward secret data entrusted to a computa-
tion to a remote principal or to a place that another untrusted module can
read
— Abuse of system resources: Consume RAM memory and disk space as
well as CPU utilization
 292

Remote Servers

Application j

Process
Applications
Module 1

Module 2

Service i
Process —'
Services Module 3

Module 2

Kernel Trusted Computing Base

Fig. 1. System Execution Model: Services and applications cire composed of pro-
cesses containing one or more dynamically-loaded modules

— Plant viruses and Trojan horses: Add binaries to the host's system and
modify the host's environment, so these malicious binaries are run by more
privileged users

The first four types of attacks are what access control infrastructure is typ-
ically designed to prevent. For example, the second type of attack can occur if
users do not adequately control discretionary access to their own executables
(e.g., a program under development). A module could identify that a file is a bi-
nary executable, and infect the program with a virus or replace it with a Trojan
horse.
The fifth type of attack is a denial-of-service of system resources. For example,
a module that writes to the disk until it is full would deny other modules from
writing to the disk. Operating systems typically do not prevent these attacks
from occurring. However, it has been shown that proper management of system
resources can enable control of the amount of service resources that a module
may consume [30]. However, these mechanisms have not been demonstrated in
practice, and it is non-trivial to determine reasonable resource limits.
In the sixth type of attack, the module leaves a trap that an unsuspecting
user may be lured into. For example, a Trojan horse program may be written
into a user's directory ostensibly used for application files, and the untrusted
module may modify the path environment (e.g., due to lax user administration
of their configuration files) resulting in the execution of the planted program
with the user's full rights.
 293

2.2 Access Control Requirements

To prevent the attacks described above, access control policy must enforce the
following constraints (at a minimum):

— Authorized object creation: Objects may be created only with authorized

names and with authorized permissions associated with them (e.g., limited
delegation, see below). For example, few, if any, principals should be able to
create executable files.
— Immutable object identities: Object names should be unique and im-
mutable to prevent time-of-check-to-time-of-use (TOCTTOU) attacks [6].
— Authorized read: Only authorized objects may be read.
— Authorized modification: Only authorized objects may be modified.
— Authorized execution: Only modules that are trusted not to extend a
principal's permissions may be executed by that principal.
— Authorized transmission: A principal may only communicate with au-
thorized principals regardless of the available communication channel (e.g.,
sockets, files, etc.).
— Authorized resource utilization: Only authorized system resources may
be used by these modules, and these resources must be revocable.
— Authorized delegation: A principal may be restricted in the rights that it
may receive via delegation. For example, it cannot delegate a principal the
right to communicate with it if a policy of separation exists between the two
principals.
— Maintain consistency with policy: The system must be flexible enough
to adjust to changes in permissions due to dynamic security policy changes.

The security requirements described above restrict the actions that a mod-
ule may perform on any system or application objects. Fundamental to these
requirements is the ability to create and name objects. In order for the system's
access control policy to be enforced on dynamically composed services and appli-
cations, it must be possible for this policy to be mapped to dynamically-created
object name spaces. These object names must be immutable to prevent renaming
after authorization (time-of-check-to-time-of-use or TOCTTOU attacks). Then,
basic access control on each operation on each object can prevent unauthorized
operations from being performed. In addition, control of resource consumption
may be desired to control some denial-of-service attacks. Next, it may be nec-
essary to control the delegation of rights by dynamically loaded modules. For
example, an application developer may not know that the delegation of a right
may lead to the creation of an unauthorized communication channel. Lastly,
this policy for controlling this delegation is determined outside the module, so it
must be possible for the system to enforce its policy as it changes (e.g., due to a
change in application state). Even though the system cannot prevent a principal
who possesses a right (in this case, the application) from using it to circumvent
security policy, it is often the case that the application wants to enforce security
policy properly, but it cannot be completely aware of that policy. The system
must be able to enforce its policy under these trust conditions.
 294

Access control models must enable the specification of the desired security
pohcies with a reasonable amount of effort. The key issue is that policies may
be dependent on dynamic factors, such as application state. For example, the
patient data available to a nursing application depends on the hospital, ward,
etc. in which the nurse principal works. In these cases, the access control model
must enable the association of application context and principals. Because policy
may be context-specific, it may be desirable for both system administrators and
other principals, such as application designers, to specify the rights available to a
module. Thus, the access control model must enable the specification of policies
by multiple principals, but these principals may not be completely trusted to
administer rights, so permission management must be restricted based on a
mandatory access control policy.

2.3 Authorization Mechanism Requirements

Once the security policy has been specified, the system must be able to enforce
it. The system must be able to derive legal permissions for its principals and
authorize their actions using these permissions.
A model for the derivation of a module's legal permissions is needed. As
described above, this model must support the abihty for multiple principals to
delegate a limited set of rights. The set of rights that may be delegated may
depend on the current state of the system or application. Therefore, there must
a mechanism for keeping the access control policy of a module consistent with
the factors it depends on (e.g., the state of the application in which it is being
run). A problem is that such management itself can become ad hoc. The model
needs a means for expressing when a modules permissions should be expanded
or retracted within the mandatory limits.
Once the a module's permissions are determined, the authorization mecha-
nism must be able to enforce these permissions. In general, authorization mech-
anisms must perform the following tasks [1]:

- Intercept all controlled operations

- Authorize each operation using the system's security requirements
- Protect the system from tampering

While the basic functions of the security mechanism remain the same as dy-
namically downloaded modules are introduced, the complexity of these tasks is
increased. First, modules may be loaded into the same protection domain (i.e.,
address space) to improve performance. However, operations within a protection
domain cannot be intercepted, so it is necessary for the security mechanisms to
determine whether such a load is permissible. Next, a dynamically downloaded
module may provide access to a resource that requires access control. For ex-
ample, the file system may be a dynamically downloaded module. However, this
module must be able to enforce system security constraints on the accesses to
its objects. Therefore, the system security mechanisms and the module need to
work together to enforce the system's access control policies properly.
 295

3 Access Control Policies

In this section, we evaluate access control modeling concepts for their ability to
describe the security policies outlined above. The requirements of these access
control models are that they be able to express the security policy requirements
and, as much as possible, ease the task of security policy specification. Role-
based access control (RBAC) models are designed to ease the management of
access control policy, so we center the discussion around these models. However,
there is no standard RBAC model, but rather, several models containing various
combinations of access control modeling features. We, therefore, examine these
various features to determine the access control policy issues they address.

3.1 Role-based Access Control

Many systems associate principals directly with users or services. A user executes
a program, and the program assumes the rights of the user principal. Unfortu-
nately, when a user executes a downloaded module, the module cannot assume
the full rights of the user without opening the user to the attacks described
above.
In general, the permissions associated with a user executing a downloaded
module would be those of that user assuming a role of executing that module. For
example, Jaeger and Prakash show that the security policy in many collaborative
situations is the intersection of the permissions of the collaborators [22]. In these
cases, a user's permissions are those of the user acting on behalf of a collaborator.
This solution is specific to collaborative systems, but a more general approach
has been known for many years: role-based access control (RBAC) [41]. In a basic
RBAC model, called RBAC96 [38], the users are associated with the roles that
they may assume. Permissions are then associated with the roles rather than
the users. This enables users to execute programs with different (hopefully, least
privilege [37]) access rights.
Many RBAC models utilize inheritance to ease the specification of access
rights. For example, users may assume application-specific roles which have a
subset of their rights as shown in Figure 2. In general, a role hierarchy represents
a set-subset relation where the rights are inherited up the tree (which is the
opposite of object-oriented inheritance). Thus, any descendant role has a subset
of its ancestor's permissions.
Thus, roles can be created with the proper rights for each module, and users
can switch to the appropriate role when they want to execute a module. In
addition, RBAC models also prescribe the use of constraints to enforce more
complex security requirements, such as separation of duty. For example, the roles
of two modules can be evaluated to show that they share no access to common
objects. Also, Chinese Wall requirements can be enforced to prevent a user from
executing two mutually exclusive sets of privileges simultaneously [39]. Even
denial-of-service requirements, such as memory usage limits, can be expressed
as constraints on the use of permissions.
 296

Root Role
RlooT U PA U ft
U R; UR) U PE

Role A RoleB
PA U B: U R] ft u R) u Pt

RoleC Role D RoleE

Ft Pb ft

Fig. 2. Role Hierarchy: Note that permissions are iniierited up the role hierarchy

TheoreticaJly, RBAC models enable enforcement of a wide variety of security

requirements and are intended to ease the task of security policy administra-
tion, but the specification of module access rights presents problems for the
basic RBAC models. First, each module may define its own objects and opera-
tions, so we examine how some additional aggregations enable role specification.
Second, a module's permissions may depend on a variety of factors, including its
application, role in the application, and application state, so we discuss a gen-
eralization of the role concept, called parameterized roles that enables a single
role definition to be applied to multiple contexts. Third, since any principal may
compose an application or service from a set of modules, the number of potential
administrators is increased, so we examine concepts that can describe role ad-
ministration. Lastly, the use of parameterized roles and multiple administrators
raises the importance of constraint enforcement, but this is not a well-developed
area of RBAC. In the following subsections, we discuss these avenues for reducing
the complexity of RBAC specification while still enabling the desired flexibility.

3.2 Aggregations

As we saw, RBAC models enable users to be aggregated into logical roles and
permissions to be aggregated for use by those roles. However, access control mod-
els can support a number of other aggregations that may ease the specification
of security policy. We examine the aggregations of objects and operations. Then,
we propose an extended RBAC model in which these aggregations are included
and assess what this extended model enables.
A number of systems enable objects with the same rights to be aggregated [19,
24,7,2,49]. In the CORBA security model [19,25], objects are aggregated into
sets called domains. Permissions are specified per domain, and any principal
can be given rights to access objects in a domain. That is, the assumption is
that different principals have common object aggregations for specifying their
rights. Of course, this may not always be the CclSGj SlS different principals may
 297

Constraints

Users

Object Operation
Groups Groups

Objects ; Operations

Fig. 3. Extended Role-based Access Control Model: Permissions are derived

from object and operation groups

use their own objects to execute may have common rights over different object
aggregations. In this case, multiple domains must be created.
Next, the operations of a specific type can be aggregated into a set of opera-
tions named by a type with common access control requirements. For example,
certain operations may be designated as read operations and others as write
operations for particular objects. This enables system administrators to specify
the rights of principals to read and/or modify data given a definition of which
operations perform those types of operations. In our work [23], application de-
velopers aggregate operations into operation groups, and system administrators
define rights in terms of those groups. Similarly, CORBA enables the definition
of interfaces, and administrators define the required rights needed to access those
interfaces. For example, a principal needs read permission on a domain to access
their read interface.
Therefore, the RBAC model presented in the previous subsection can be ex-
tended as shown in Figure 3. In this RBAC model, objects for which common
rights may be expressed are aggregated into object groups. Operations which cor-
respond to common operation types can be aggregated into operation groups.
Then, the set of permissions is a relation between the sets of object groups and
operation groups. Further, object groups and operation groups can themselves
be aggregated, and these aggregations may be subject to constraints. For ex-
ample, an object group can be created that contains the intersection of two
object groups' members. Also, permissions can be constrained based on their
constituent object and operation groups. For example, this would enable the
specification of the intersection of two permission sets.
An obvious benefit of this extended RBAC model is that the task of express-
ing permissions is eased because it is possible to express permissions using fewer
statements. An additional, less obvious benefit is that it enables a separation
of labor between users/developers and system administrators. Because system
administrators often do not know the semantics of application objects and op-
 298

erations, they cannot effectively express rights for them. Instead, apphcation
developers may create object groups and operation groups that map to semanti-
cally meaningful system administrator notions. While this may seem dangerous,
it is generally not unreasonable for modules to define the access rights of others
to their objects.
Unfortunately, this does not solve all our problems and creates some new
ones. Object and operation aggregations do not provide the flexibility needed in
some cases because they are static for all execution instances of an application.
For example, in a collaborative context, only the objects in use in that con-
text should be accessible to the collaborators. Therefore, an object aggregation
should depend on the application state. Also, a new problem is created because
application developers and users can assign rights to other principals. Every right
that is delegated to a principal should be within the system's security policy.
This does not preclude us from establishing that modules can enforce security
policy on their objects - we must trust them to provide proper access to the ob-
jects they serve. However, module writers cannot be trusted to know the system
security policy, so they cannot be allowed to delegate permissions arbitrarily.
Constraints, such as separation of duty, may be violated. In the following two
subsections, we address these two problems.

3.3 Context-Sensitive Roles

Context-sensitive RBAC models enable permissions to be specified relative to a

context [15,31]. Giuri and Iglio formahzed this notion in terms oi parameterized
roles. A parameterized role consists of a role identifier and a set of parame-
ters whose values determine the context. Object groups in the permissions are
indexed by the parameters to indicate a particular set of objects.
Thus, by setting different parameter values, different permissions can be ob-
tained. For collaborative applications, we use parameterized roles to represent
the permissions of each collaborator [23]. Each collaborator is assigned a role
that is parameterized by the user principal, module author (i.e., collaborator),
application, module's role in the apphcation, and the application's state. Thus,
the rights associated with the role are dependent on the user's objects that are
being used in this collaborative session. Therefore, least privilege rights are en-
forced on the collaborators because only the objects in use in this instance of
the collaboration are accessible.
Context-sensitive roles can represent a set of roles using only one role spec-
ification. We also use context-sensitive roles to enable system administrators to
specify the permissions which may be assigned to an application for any user,
and for application developers to specify permissions for each execution of their
applications. In this model, many principals may be able to specify roles, so
the need for proper role administration becomes much more important. Also,
the specification of roles must adhere to the system's security policy. Since con-
straints are used to enforce system security policy, the ability to specify and
enforce constraints becomes more important.
 299

Fig. 4. Transform Limits: A principal's permissions are at most the union of the
permissions that may be delegated by principals 1,2, and 3.

Since a principal may obtain permissions based on a change in context, it

may be necessary to restrict the permissions that may be delegated to adhere to
system and application security policies. We envision that services will be com-
posed from a set of modules where one distinguished module may be responsible
for the correct behavior of the service. For example, consider a virtual mem-
ory system composed from several modules. The main virtual memory module
enables the download of application-specific memory managers on the behalf
of applications. Since memory manager modules may be requested to transfer
pages among one another (e.g., to share memory), the virtual memory system
must restrict this with respect to the system's security policy (or else it does
not behave correctly. Thus, we make a distinction between the permissions that
a principal possesses and a principal's transform limits [23]. A transform limit
defines the set of permissions that a delegator may grant to a delegates. The
union of a principal's transform limits determine the maximal permissions that
that principal may obtain (see Figure 4). In this example, the virtual memory
system can restrict the permissions that its modules can distribute among one
another. These restrictions depend on the context in which the module is used,
so context-sensitive roles are necessary to describe such restrictions.

3.4 Administration
RBAC models have also been designed that support the administration of se-
curity policies. In the ARBAC 97 model [40], an administrative role hierar-
chy defines the administrative roles and their rights to modify the information
in the basic role hierarchy: user-role assignment, permission-role assignment,
and role-role assignment (i.e., construction of the role hierarchy itself). Based
on the extensions described above, ARBAC 97 could likewise be extended to
enforce object-object group assignment, operation-operation group assignment,
and object-operation assignment (i.e., creation of permissions).
 300

The ARBAC 97 model is a general model for expressing administrative roles

and permissible duties. The authors typically use an example in which admin-
istrative roles are associated with organizational groups. No explicit semantics
are associated with these administrative roles, but it is likely that the intent
is that the administrative roles control roles, objects, and permissions in their
organizational group. However, we are not aware of any use of administrative
role hierarchies, so the construction of effective administrative hierarchies is aji
open issue.
Some work has considered the role of application developers and users in
administration of roles. The NAPOLEAN system divides access control admin-
istration into seven layers [43]. The application developers are responsible for
defining the bottom four layers: objects, object handles (operations), applica-
tion constraints, and application keys (actually application roles). System ad-
ministrators define the top three layers: enterprise constraints, enterprise keys
(system roles), and key (role) chains. Despite the non-standard nomenclature,
a number of important RBAC features are represented in NAPOLEAN: (1) ap-
plication developers can define the object name spaces and application-specific
roles and (2) system administrators associate users with these roles (or their
aggregates, key chains) and restrict the rights that principals may be granted
(using constraints). The enterprise constraints are not described in any detail,
so we suspect that they are ad hoc.
In our access control model [23], role administrators (i.e., administrators of
the role as specified by the role administration hierarchy) specify the transform
limits of the role. There is a distinction made between the administrators of the
role, who are trusted to grant all of their rights to the role, and the delegators
of role, who are restricted in the rights that they may delegate by the transform
limits. This distinction enables the role administrator to specify limits on the
transfer of rights, so that their access control policy can be enforced. Such policies
may be dynamic, however, so simply restricting the possible delegations may not
be sufficient to enforce them. Consider the Chinese Wall policy [8] where one
of disjoint subset of permissions is obtained based on an action. In this
single set of transform limits for a principal is insufficient because the transform
limits that need to be enforced depend on which operations the principal has
performed. It is envisioned that role constraints must be able to adjust the
transform limits based on the current policy.

3.5 Constraints
The use of constraints in RBAC is the least-developed notion in RBAC models.
Currently, research has identified two sub-satisfactory notions of constraints. In
some systems, a general notion of constraints is included in the model, as is done
in ARBAC 97 or NAPOLEAN, that does not include a language for specifying
such constraints. In other systems, specific, limited constraints are supported.
However, these constraint languages are not general purpose.
A security constraint language must enable enforcement of the variety of
security constraints. For example, it must be possible to specify restrictions
 301

on the rights a principal may attain, that principals cannot modify executables,
separation of duty between principals, Chinese Wall constraints, denial-of-service
constraints, etc. However, some constraints are distinctly different, so entirely
different constraint models may be appropriate. For example, enforcement of
the Bell-LaPadula security policy is much different than the basic access control
policy outlined here. While it has been shown that general RBAC models can
enforce Bell-LaPadula [3], it is not clear whether a general constraint language
should be exposed to the administrators who write the policies.
In our work, we take a two-pronged approach. First, we enable the speci-
fication of transform limits. A transform limit specifies the set of permissions
that a delegatee can obtain from a specific delegator. The set of transform limits
determine the maximal set of permissions that a principal can obtain. Second,
the definition of such transform limits depend higher-level security constraints,
such as Chinese Wall [8]. A Chinese Wall policy is a dynamic separation of duty
policy in which only one of several disjoint sets of rights may be used depending
on the actions of the principal. Constraints are written that determine the trans-
form limits that may be active. The language for expressing such constraints is
still under development.
Bertino et al propose a language for defining constraints on the assignment
of users to roles and the creation of inheritance role hierarchies [5]. Their ap-
proach is designed to enforce constraints on the execution of workfiow tasks. For
example, a user may be restricted from executing a second task if he executes a
previous task in a particular role. Therefore, constraints are specified in terms
of tasks. Constraints specific to operations within a task cannot be expressed,
however.
Our opinion is that a single general constraint language should be devised.
This language should be general enough to express any constraint (on users, ob-
jects, operations, and their administration). However, such a language will likely
be too abstract and complex for use by system administrators, so application
specification languages must be built on this base constraint language for the
individual applications.

4 Access Control Enforcement

In general, access control policy enforcement requires two mechanisms: (1) a

mechanism to derive the access control policy associated with each principal
(permission management mechanism) and (2) a mechanism to authorize con-
trolled operations using this access control policy (authorization mechanism).
We examine the functionality offered by traditional implementations of these
two mechanisms and discuss how they must be altered to support our require-
ments. In the next section, we describe how the mechanisms provided in the
Lava Security Architecture address these requirements.
 302

4.1 Permission Management

Permission management mechanisms determine each principal's access rights at

any time. Lampson has formahzed the representation of access rights into an
access matrix whose two practical variations are access control lists and capabil-
ities [27]. An access control list associates an object with the principals and the
operations they can perform on the object. A capability defines objects and the
operations that can be performed upon them by any principal that holds the
capability. Since our access control model involves the dynamic modification of
principal rights, a capability approach is easier to maintain.
The early capability-based systems took the view that each invocation of a
procedure may have its own access rights. Dennis and Van Horn formalized this
notion into a concept called a protected procedure call [10]. A protected procedure
call is designed to mutually protect the caller and the procedure from each other.
The capabilities available to the procedure are the union of its own capabilities
and the capabilities that the caller explicitly passed in the invocation. Thus, the
procedure only gains access to the capabilities that the caller grants, and the
caller does not obtain direct access to the capabilities of the procedure.
While many systems used the protected procedure call approach (e.g., [9,
48]), other system designers found its semantics of permission management in-
adequate. For example, the CAP system [47] derives a procedure's capabilities
from a variety of inputs which, in addition to those of the procedure and those
passed by the caller, include: (1) the capabilities of the process in which the pro-
cedure is run; and (2) the capabilities of all system procedures. Mungi protection
domain extension enables capabilities not explicitly passed in the procedure call
to also be provided to the procedure, but only for the duration of the procedure
call [44].
Problems of confinement and revocation have made capability-based systems
difficult to use to enforce system security policies. In general, the possession of
a capability is sufficient to use it. Since a capability can be distributed in a
procedure call, it is hard to prevent principals from being given unauthorized
capabilities and it is hard to revoke capabilities once they have been granted.
Enforcement of permission limits, such as transform limits, enable the system to
restrict the set of capabilities that a process may obtain. Approaches that enable
revocation involve some form of indirection. Redell's revoker capabilities enable
invahdation of delegated capabilities. DTOS [33], MOSES [34], and Lava [21]
cache capabilities in a trusted process, so immediate revocation is possible.
JDK 1.2 uses an approach to permission management that enables permis-
sions to be determined based on run time factors, called stack introspection [17,
45]. Using stack introspection, the permissions of an executing thread are the
intersection of the permissions of the classes whose methods have been executed.
A class's permissions restrict the set of capabilities that the class's methods may
ever obtain. In JDK 1.2, the permissions of a class depend on the class's provider
and/or code base. In order for the class to execute properly in all reasonable con-
texts, the rights assigned to a class must be the union of all the rights that the
class may ever need. Therefore, the intersection of these rights may still grant
 303

r •\
>
Module Process
Security Manager
Transforms Using
Transform Principal 1
Limits ^^.^r Permissions

Operation
Principal 2
Permissions
\ ^.
Principal 3
Permissions

Fig. 5. Transforms: A module's operations may be associated with transforms that

grant rights to or revoke rights to one or more principals, perhaps including the mod-
ule's principal itself.

the thread many rights that are not necessary given the appHcation in which the
class is used, the purpose of the class in that application, and the application's
current state.
Instead, we advocate an approach where a principal's permissions are always
consistent with the current state of the application or service. Using our access
control model, these delegations are restricted by the transform limits of the
principal. The problem is to automate the process of delegation. For this purpose,
we define the concept of a transform which associates an operation with a change
in permissions of one or more principals [23] (see Figure 5). Thus, changes in
application state can be correlated to changes in permissions. These permissions
changes are authorized using the transform limits of the delegatees.
Once a principal obtains a permission, a capability can be obtained for it. For
example, possession of a permission to read and write a file enables the principal
to obtain a capability with those operations activated. When a transform revokes
a principal's permission, the system may need to revoke some of its capabilities.
However, multiple principals may grant the same permission, so any effected
capabilities are only invalidated. They are authorized again upon the next use.

4.2 Authorization Mechanisms

Authorization mechanisms determine whether a principal is permitted to invoke

an operation on an object. Logically, an authorizing agent retrieves the princi-
pal's access control policy and compares the request (perform an operation on
an object) to the policy. If an entry in the policy grants the request and no policy
precludes that request, then the authorizing agent permits the request.
Currently, most operating systems (e.g., UNIX and Windows NT) are mono-
lithic, and they only enforce access control on system calls. Since the kernels
store all objects to which access is controlled, they have access to the name
spaces of the objects they need to protect (e.g., the file system). Therefore, the
 304

kernel uses its internal representation of the system name space and its internal
representation of the system security policy to authorize system calls.
There are three problems with this approach for systems configured from
modules dynamically: (1) the name spaces of objects that need to be controlled
may be outside the control of the kernel; (2) the kernel does not aid the applica-
tions in enforcing their security requirements; and (3) the access control policy
flexibility of our model increases the kernel's complexity. First, modules manage
their own name spaces, but these name spaces must be predictable, so system
and application security policies can be enforced. Second, downloaded modules
may want the system's trusted computing base to help it enforce its access con-
trol requirements. For example, an application may use another module, but this
module should only have a subset of the application's rights. Lastly, the addition
of more flexibility to security policy means that the authorization mechanisms
become more complex. In a configurable system, inter-process communication
(IPC) must be very fast to support a reasonably fine granularity of processes,
but the addition of a complex authorization mechanism into each IPC would
add a mandatory cost to all IPCs (even those with few requirements to check).
Micro-kernel systems (e.g., Mach [36] and Spring [13]) avoid the name space
problem by only controlling communication between processes. For a client pro-
cess to invoke an operation on a server process, the client must have a capability
to send a message to a port to which that server has a receive capability. The
semantics of the operation are determined by the server and authorized by the
server in its own way. Also, capabilities can be passed freely in messages, so
no control on delegation is enforced. While this removes the complexity of the
name space from the kernel, it prevents the kernel from being able to enforce
the system's current security policy and requires each server to define its own
authorization mechanism which may result in errors.
Reference monitors enable control of delegation. For example, the Distributed
Trusted Operating System (DTOS) extends the Mach kernel by a mechanism
that enables custom security policies to be enforced by security servers [33].
DTOS adds a capability cache to the kernel, so servers can call the kernel to
authorize operations. If a capability is not present, a security fault is taken which
results in a call to the appropriate security server. Security servers interpret the
server's object name spaces. The administrators of the service that use this se-
curity server must know enough about the object name spaces of these servers to
enforce the policy. Upon a successful authorization, the security server then up-
dates the capability cache. Also, security servers can invalidate the entries in the
capability cache which enables revocation. DTOS provides a flexible authoriza-
tion mechanism, but it makes the kernel more complex (which has been found
to degrade IPC performance [29]) and requires extra communication overhead
to authorize the operation (1 round-trip IPC).
Our goal is to combine the kernel cache and security servers into one entity
outside the kernel. The Clans & Chiefs mechanism is not explicitly an authoriza-
tion mechanism, but it can be applied to authorization [28]. In this mechanism,
a chief is a special process in a clan (i.e., set of processes). Any IPC between
 305

two processes in the same clan are forwarded directly to the destination by the
kernel. However, any IPC either to a process outside the clan or from a process
outside the clan is automatically redirected by the kernel to the chief. In a secu-
rity scenario, the chief can authorize any communication between processes in
its clan and the other processes in the system. In an early version of the Lava
Security Architecture [21], we use a chief for every dynamically loaded module,
so we could control its accesses to the trusted servers and other dynamically
downloaded modules. The chief stores the access control policy of the process
that it controls and authorizes all of this process's operations. Of course, if an
IPC is between two different clans, then three system IPCs (client-chief, chief-
chief, chief-server) are required for the communication to be complete. While we
showed that the base cost could be as low as 4 /xs and that we measured 9 fis,
the extra IPC (chief-chief) performed no useful operations.
The problems with the Clans & Chiefs mechanism are: (1) that every process
in a clan must be able to freely communicate with every other process in the
clan and (2) at least two chiefs must be invoked for every inter-clan IPC. In
order for two processes to belong to the same clan they must be able to freely
communicate, freely distribute capabilities, and avoid revocation of the other's
capabilities for their entire execution. To interpose a chief requires that the pro-
cesses be deleted and re-created in the desired clans (i.e., clan relationships are
static). Therefore, most of our clans have degenerated to a single clan member.
Also, it has been shown that a single reference monitor is capable of enforcing
very complex security requirements. The need to go through multiple chiefs, by
default, is too expensive.

Therefore, we evolve the Clans & Chiefs mechanism to better satisfy these
goals. In general, the notion of the IPC redirection is powerful: an IPC can either
be sent to the destination process or redirected to another process. Therefore, we
endow the kernel with the ability to maintain a mapping between destinations
and redirections for each process [20]. For example, consider Figure 6. When the
source sends an IPC to a destination, the kernel is invoked. The kernel examines
the redirection cache to determine if there exists an entry for the destination. For
example, a particular reference monitor may be assigned to this communication
channel. In this case, a reference monitor receives the redirected IPC and can
authorize it before sending it to the destination.
An effective reference monitor must be able to mediate all communication,
but how the mediation is done can be controlled flexibly using this mechanism.
In general, a reference monitor between a particular source and destination ver-
ifies that the source has the ability to invoke the requested operation on the
destination (i.e., within the source's current permissions). Thus, the system's
security policy can be enforced on each IPC. If the destination is trusted to
enforce the system's security policy on a particular source, then the reference
monitor may be removed from that communication channel (i.e., the redirec-
tion is changed from the reference monitor to the server). However, if the access
control policy should change, then the IPC redirection mechanism enables the
 306

Source Destination

Redirection

Kernel
Redirection Cache
Destinations

Fig. 6. IPC Redirection: An IPC from a source to a destination may be redirected

if an entry is present in the redirection cache.

system to reactivate the reference monitor for this communication channel, so

the more restrictive security requirements can be enforced.
A problem is to decide who is able to set the redirection policy for a process.
We propose that this is where the notion of a chief appears. To avoid confu-
sion, we rename the chief to redirection coordinator. A redirection coordinator
is permitted to set the redirection entries for the processes that belong to its
redirection set. Therefore, arbitrary configurations of reference monitors can be
constructed dynamically to control the processes in a redirection set.

5 Lava Security Architecture

The Lava Security Architecture uses a role-based access control (RBAC) model
to represent system access control policies and general mechanisms for permis-
sion management and authorization to enforce these policies. The architecture
enables each module to be assigned access rights based on their purpose in an
application, and to enforce those rights with a minimal communication overhead
depending on the security policy. In this section, we outline how the architecture
can achieve these goals.
The Lava Security Architecture consists of a system nucleus, security archi-
tecture interface (SAI) module, a set of reference monitors, a transform Ubrary,
 307

Client Server '

Reference
Transform Monitor Authorization Transform
Module Library Library Library
, ^ )
Security Arcliitecture Interface

Access Control Policy

(from a secure server)

Lava Nucleus
Redirection
Cache

Fig. 7. Lava Security Architecture: Consists of the following components: (1) a

Lava nucleus that provides fundamental 0/S primitives and redirected IPC; (2) a
Security Architecture Interface that enables modules to be loaded and the system access
control policy to be enforced; (3) reference monitors that control module permissions;
(3) an authorization library that manages and authorizes the use of server capabilities
relative to system policy; and (5) a transform library that supports the execution of
delegations, so permissions can be maintain consistently with application state.

and an authorization library (see Figure 7). The Lava nucleus provides operat-
ing system primitives upon which all applications and services are constructed:
multi-threaded tasks, address spaces, and inter-process communication (IPC).
The nucleus uses the system's hardware protection to separate individual tasks
and IPC redirection to control communication. The SAI derives the access con-
trol policy for each module and loads the modules into tasks in such a way that
this policy can be enforced. The SAI assigns reference monitors to enforce a
task's access control policy. A reference monitor can mediate all IPCs on the
communication links it has been assigned, so it can restrict communication be-
tween any two tasks, restrict delegation of permissions, restrict the operations its
tasks can perform, and revoke permissions if the access control policy changes.
The transform library supports the execution of transforms upon module oper-
ations, so that the access control policy can be maintained consistently with the
application's state. The authorization library is provided to servers, so they may
enforce system security policy on their objects without the need to build ad hoc
security infrastructures.
The access control policy data must be distributed carefully between the
servers and reference monitors to enable flexible and secure authorization as
shown in Figure 8. Each reference monitor maintains its task's permissions.
Permissions are stored per server in server permission tables. The reasons storing
permissions per server are: (1) servers must be able to authorize the creation
of capabilities to their objects, but we do not want them to have access to
permissions for another server and (2) this reduces the memory fragmentation
 308

Reference Monitor Server

Server Capabilities Table

(may use server's copy, read-only)

Valid References Bitmap

Read-Only

Fig. 8. Access Control Data: (1) server permissions table which holds the permis-
sions of each task organized per server, so they may be mapped read-only to the server
for authorizing capability creation; (2) server capabilities table which contains the ca-
pabilities the server has created for each task; and (3) valid references table which tells
which server capabilities have been authorized for use by which tasks.

that w^ould occur if the permissions were stored per task and server. However,
the result is that all reference monitors have access to the same server permission
table pages, so some concurrency control on update is required. Note that the
servers have read-only copies of their permissions, so the reference monitors have
control over what the system policy is.
Each server and reference monitor may maintain their own server capabilities
table in which capabilities to the server's objects are stored. Obviously, the refer-
ence monitor's copy of a server's capabilities is redundant, but may be necessary
because a server may not be trusted to maintain its own capability information
securely (i.e., an authorization mechanism must ensure that the security data is
tamperproof). For trusted servers, a single copy of its capabilities may be stored
and used by reference monitors as well (read-only). When a capability is created,
the servers return capability references to the holder of the capability. This refer-
ence refers to the index of the capability in the server's server capabilities table.
The reference monitors may obtain read-only access to the server capabilities
table which they use to authorize delegations and revoke capabilities to objects
when permissions are removed.

The reference monitors must be able to revoke a capability reference at any

time. For example, the principal's policy may change resulting in a removal
of some of its permissions. The reference monitor must be able to revoke the
associated capabilities immediately, but it cannot modify the server capabilities
table. Instead, the reference monitors maintain valid references bitmaps which
indicate which capability references are valid for a task.
 309

5.1 Loading Modules

The security architecture interface (SAI) enables modules to be loaded into Lava
tasks (see [23] for a detailed discussion). A task may request that a module
be loaded into its own or a new task. Depending on the loading requirements
(the specified task, application, and/or application role) the SAI retrieves the
module's access control requirements from a secure server. The access control
policy is represented using context-sensitive roles where the context parameters
are defined to be: (1) the downloading principal; (2) the module's author; (3) the
apphcation in which the module is to be used; (4) the role the module assumes
in the application; and (5) the specific application instance. Using this identity,
the SAI derives the transform limits for the module. Thus, it is possible for a
module to be a specific type of extension (i.e., role) to an instance of a system
service or application. The module is then loaded into its task and transform
limits are assigned to reference monitor for enforcement. If a module is loaded
into an existing task the transform limits of that task are the intersected.
Also, module developers may specify information to keep the access control
policy consistent with the application's state. First, they may specify mappings
between their operations and known operation aggregations, such as read and
write. Thus, module operations can be mapped to system security policy re-
quirements. Also, module developers may specify transforms that describe how
module operations trigger the delegation of permissions to individual delegatees.
The SAI is the redirection coordinator for the system. It assigns reference
monitors to tasks by setting the redirection entries for those tasks. The Lava
nucleus provides a system call that enables the SAI to set any redirection entry
for any task in its redirection group (in this case, all tasks created by the SAI).
The Lava nucleus then uses the redirection entries to forward the IPC to the
task specified in the redirection entry.
Because the nucleus IPC mechanism performs redirection, the size of the
redirection data might become important. If the amount of redirection data is
large, then it is more likely that data cache misses will occur on examiniation
of redirection data. In general, a redirection entry is needed for each possible
combination of sender task and receiver task. Thus, the number of potential
redirection entries grows by the square of the number of system tasks. In many
cases, optimizations are possible. For example, we can assign a reference mon-
itor to each task, but permit some exceptions to be created for which no or
different redirections are possible. Therefore, we only need to store a task's redi-
rection default (i.e., its reference monitor) and the exceptions. If the number
of exception is small, then the data needed will be small. For some tasks this
optimization may not apply, of course, so we are actively investigating useful
compact representations.

5.2 Managing Permissions

The Lava Security Architecture implements the following model of permission
management. Each task is assigned to a principal, and each principal has a ref-
erence monitor that stores its set of transform limits and current permissions.
 310

A principal's current permissions at any time must be a subset of its transform

limits. A principal may delegate permissions to another principal if the dele-
gatee's transform limits allow it. Such delegations are automated by defining
a transform that specifies a set of rights to be delegated and/or revoked. The
module developer associates the transforms with the operations that trigger the
change in rights.
Each task is provided with a transform library that enables transforms to
be executed. The transform library forwards the transform to the delegatee's
reference monitor. The reference monitor authorizes the transform using the
delegatee's transform limits. If authorized, the reference monitor updates its
current permissions for the principal. Multiple principals may delegate the same
permissions independently, so each permission is associated with its delegator.
A task can obtain capabilities for any permission that it possesses. A capa-
bility consists of the following fields: (1) server id; (2) object type; (3) module
instance id; (4) object id; (5) operations; and (6) server-specific data. Modules
may define multiple interfaces, and each interface is an object type. Multiple in-
stances of the same module may be loaded into a task, so the module instance id
differentiates them. The server-specific data is used to express server-specific con-
straints, such as those pertaining to denial-of-service. To prevent modification,
prevent forgery, and be able to revoke capabilities, the object id field actually
stores an index to the capability in the server capability table (i.e., the capability
reference).

5.3 Authorizing Operations

The reference monitors and the authorization library work together to authorize
operations. An important goal is that the architecture enable the minimal au-
thorization overhead given the security requirements that need to be enforced.
Thus, the system supports the four monitoring configurations shown in Figure 9.
In the first case, each task has a reference monitor that authorizes its operations.
In the second case, the system's access control requirements permit direct com-
munication between the client task and the server task. The third and fourth
cases monitor the IPCs sent from one of the two tasks.
We now examine the authorization mechanism for case one. In this case, the
task's operation request is intercepted by its reference monitor (step 1). The
reference monitor uses the valid references bitmap for the task and server to
determine if there is a vahd capability that corresponds to that reference. There
are three possibilities: (1) the reference is not valid, but there is a corresponding
capability; (2) the reference is not valid, and there is no corresponding capability;
and (3) the reference is valid. If the capability reference is not valid, the reference
monitor determines if the capability associated with that reference is authorized
by the task's permissions. If there is no capability associated with the reference,
then the reference monitor takes a reference fault and obtains the reference's
corresponding capability from the server. The reference monitor then authorizes
the capability using the task's server permissions. If authorized, the capability
is marked valid and the reference monitor allows the request. Note that if the
 311

Case One Case Twc>

RM \2 4/ RM
2

•/ X
r N
\ ' Task
f 1
^
\
Server
Task Server

Case Three Case Four

RM \2 RM
3 /

•/.
Task
^
3

Server Task
1 V
Server
'

Fig. 9. Authorization Configurations: Using IPC redirection these four configura-

tions of authorization can be supported on each communication channel.

server is trusted to maintain its capabilities, then no reference faults will occur
because the reference monitor will have access to these capabilities.
If the operation is authorized by the reference monitor any capability ref-
erences in the request are also passed to the server (step 2). This delegation is
authorized by the server's reference monitor when the server tries to use them.
Since the capability reference bitmap will indicate that the capability is invalid,
the server must have a permission that enables it to use that capability. The
server authorization library ensures that all capabilities for the same server ob-
ject are stored at the same reference, so it is not possible to get a false positive.
The server uses the authorization library to authorize the operation requested
using the system's security policy as well as its own. First, it checks its server
capabilities table for the principal and capability. As described above, a capabil-
ity may have been delegated to the requesting task by another task. In this case,
the server authorizes the creation of a capability using the server permissions for
the requesting teisk. The server may enforce additional security requirements on
the client task which it specifies to the authorization library.
If the server generates a capability as a result of the operation, it uses the
server authorization library to authorize the creation (using the server permis-
sions) and to add it to the server capabilities table for the task. The server
returns a capability reference which is an index to the capability in the server
capabilities table.
 312

A capability reference may be returned in a response to an operation (step 3).

As in the case of references passed in an operation, the reference monitor of the
delegator does not restrict the distribution of references (step 4). The server's
reference monitor verifies that the server is able to send responses to the task.
The task's reference monitor will authorize the reference upon its subsequent
use.
Removal of reference monitors in cases two, three, and four implies that:
(1) the task can communicate with the server and vice versa; (2) the validity of
capability references is not checked; and (3) system policy changes may not affect
current capabilities ^. In cases two and four, the task sends an operation request
directly to the server. Thus, the server alone determines whether the request
is authorized using the server permissions and capabilities as described above.
Thus, the server must track the system's server permissions and be trusted to
enforce the system's security policy. Since the server may write to its server
capabilities table, this trust implies that the server is trusted not to accidentally
modify this data. Also, a change in security policy does not change the set of
capabilities in the server. The server could maintain consistency between its
capabilities and system policy, but this cannot be enforced. Note that at any
time the reference monitor can be re-activated to mediate communication on
any communication channel, so should policy change the reference monitor can
enforce it.

6 Conclusions
A system that composes services and applications from modules that are dynam-
ically downloaded must address the possibility that the modules used may not be
fully trusted. However, these may be binary modules, so a wide variety of system
compromises are possible. The fundamental means for controlling the operations
that a module can perform is access control. Access control consists of two main
functions: (1) access control policy specification and (2) access control policy
enforcement. In this paper, we survey the representations for flexible access con-
trol policy specification and mechanisms for enforcing these policies to identify
their useful features. We then describe how the Lava Security Architecture takes
advantage of these useful features.
The main concern in the composition of applications and services from mod-
ules is that the permissions associated with a module are consistent with the
module's purpose (i.e., least privilege). Current systems associate permissions
either based on the executor of the module (e.g., in operating systems) or the
provider and name of the module (e.g., in language-based systems). However,
this information does not describe the permissions that are consistent with the
application or service's state and the role of the module in the application or
service. We develop a role-based access control model that enables permissions
' Some denial of service attacks may also be thwarted by reference monitors, so the
removal of a reference monitor also implies that the server is not susceptible to these
attack
 313

to be maintained consistently with application state. The main features of this

model are that: (1) context-sensitive principals can be defined that express a
module's permissions based on its executor, provider, application, application
role, and application state; (2) aggregations of objects and operations can be
created to express the module's context; and (3) transform limits that restrict
the rights that a principal may be granted due to context changes. Constraints
ensure that the transform limits themselves are consistent with security policy
as it evolves (e.g., due to dynamic policies, such as Chinese Wall). Work on
constraint models for RBAC is fairly immature at present.
An access control mechanism must be able to manage permissions consis-
tently with access control policy and enforce them efficiently. In current systems,
permission management does not account for application state, and most autho-
rization mechanisms do not address keeping the task permissions consistent with
the system's security pohcy. We describe how transforms can be added to mod-
ules and how the system maintains task permissions relative to application state
and within system policy. We define how a mechanism for IPC redirection en-
ables minimal overhead in the enforcement of access control policy. At most,
one reference monitor needs to be invoked to authorize communication, but, in
under some conditions, even that reference monitor is not necessary. However,
the IPC redirection mechanism enables a reference monitor to be interposed on
any communication path dynamically.

Acknowledgements

I would like to acknowledge the support of the Lava project team, particularly,
Kevin Elphinstone, Jochen Liedtke, Seva Panteleenko, and Yoon Park. Other
helpful direction has been provided by Ed Felten, Peter Honeyman, Li Gong,
Atul Prakash, Avi Rubin, Jonathan Shapiro, Leendert van Doom, and Dan
Wallach.

References

1. J. p. Anderson. Computer security technology planning study. Technical Report

ESD-TR-73-51, James P. Anderson and Co., Fort Washington, PA, USA, 1972.
2. L. Badger, D. F. Sterne, D. L. Sherman, K. M. Walker, and S. A. Haghighat.
Practical domain and type enforcement for UNIX. In IEEE Symposium on Security
and Privacy, pages 66-77, 1995.
3. D. E. Bell and L. J. LaPadula. Secure computer system: Unified exposition and
Multics interpretation. Technical Report MTR-2997, Mitre Corporation, January
1976.
4. B. N. Bershad, S. Savage, P. Pardyak, E. G. Sirer, M. E. Fiuczynski, D. Becker,
C. Chambers, and S. Eggers. Extensibility, safety, and performance in the SPIN
operating system. In Proceedings of the 15th Symposium on Operating Systems
Principles, pages 267-284, 1995.
 314

5. E. Bertino, E. Ferrari, and V. Atlttri. A flexible model for the specification and
enforcement of role-based authorizations in workflow management systems. In
Proceedings of the Second ACM Role-Based Access Control Workshop, November
1997.
6. M. Bishop and M. Dilger. Checking for race conditions in file accesses. Computing
Systems, 9(2):131-152, 1996.
7. W. E. Boebert and R. Y. Kain. A practical alternative to hierarchical integrity
policies. In Proceedings of the 8th National Computer Security Conference, pages
18-27, 1985.
8. D. F. C. Brewer and M. J. Nash. The Chinese Wall security policy. In Proceedings
of IEEE Symposium on Security and Privacy, pages 206-214, 1989.
9. J. S. Chase, H. M. Levy, M. J. Feeley, and E. D. Lazowska. Sharing and protec-
tion in a single-address-space operating system. ACM Transactions on Computer
Systems, 12(4):271-307, November 1994.
10. J. B. Dennis and E. C. Van Horn. Programming semantics for multiprogrammed
computations. Communications of the ACM, 9(3):143-155, March 1966.
11. S. Dorwaid, R. Pike, and P. Winterbottom. Inferno: la commedia interattiva, 1996.
Available from inferno.bell-labs.com.
12. D. Engler, F. Kaashoek, and J. O'Toole. Exokernel: An operating system ar-
chitecture for application level resource management. In Proceedings of the 15th
Symposium on Operating Systems Principles, pages 251-266, December 1995.
13. J. G. Mitchell et al. An overview of the Spring system. In Proceedings of Compcon,
February 1994.
14. B. Ford, G. Back, G. Benson, J. Lepreau, A. Lin, and O. Shivers. The Flux OSKit:
A substrate for kernel and language research. In Proceedings of the 16th Symposium
on Operating Systems Principles, pages 38-51, 1997.
15. L. Giuri and P. Iglio. Role templates for content-based access control. In Proceed-
ings of the Second ACM Role-Based Access Control Workshop, November 1997.
16. Y. Goldberg, M. Safran, and E. Shapiro. Active Mail - a framework for imple-
menting groupware. In CSCW 92 Proceedings, pages 75-83, 1992.
17. L. Gong. Java security: present and near future. IEEE Micro, 17(3):14-19, 1997.
18. L. Gong, M. Mueller, H. Prefullchandra, and R. Schemers. An overview of the
new security architecture in the Java Development Kit 1.2. In Proceedings of
the USENIX Symposium on Internet Technologies and Systems, pages 103-112,
December 1997.
19. Object Management Group. Security service specification. In CORBAservices:
Common Object Services Specification, chapter 15. November 1997. Available from
http://www.omg.org.
20. T. Jaeger, K. Elphinstone, J. Liedtke, V. Panteleenko, and Y. Park. Flexible access
control using IPC redirection. In Proceedings of the 7th Workshop on Hot Topics
in Operating Systems, 1999. To appear.
21. T. Jaeger, J. Liedtke, and N. Islam. Operating system protection for fine-grained
programs. In Proceedings of the 7th USENIX Security Symposium, pages 143-156,
January 1998.
22. T. Jaeger and A. PraJiash. Support for the file system security requirements of
computational e-mail systems. In Proceedings of the 2nd ACM Conference on
Computer and Communications Security, pages 1-9, 1994.
23. T. Jaeger, A. Prakash, J. Liedtke, and N. Islam. Flexible control of downloaded
executable content. ACM Transactions on Information System Security, May 1999.
To appear.
 315

24. T. Jaeger, A. Rubin, and A. Prakash. Building systems that flexibly control down-
loaded executable content. In Proceedings of the 6th USENIX Security Symposium,
pages 131-148, July 1996.
25. G. Karjoth. Authorization in CORBA security. In Proceedings of ESORICS '98,
1998.
26. M. Knister and A. Prakash. Issues in the design of a toolkit for supporting multiple
group editors. Computing Systems, 6(2):135-166, 1993.
27. B. Lampson. Protection. ACM Operating Systems Review, 8(l):18-24, January
1974.
28. J. Liedtke. Clans & chiefs. In Architektur von Rechensystemen. Springer-Verlag,
March 1992. In English.
29. J. Liedtke. Improving IPC by kernel design. In Proceedings of the 14th Symposium
on Operating Systems Principles, pages 175-187, 1993.
30. J. Liedtke, N. Islam, and T. Jaeger. Preventing denial-of-service attacks on a
/i-kernel for WebOSes. In Proceedings of the Sixth Workshop on Hot Topics in
Operating Systems, pages 73-79, May 1997.
31. E. C. Lupu and M. Sloman. Reconciling role-based management and role-based
access control. In Proceedings of the Second ACM Role-Based Access Control Work-
shop, November 1997.
32. S. D. Majewski. Distributed programming: Agentware/ componentware/ dis-
tributed objects. Available at h t t p : / / minsky.med.virginia.edu/ sdm7g/ Projects/
Python/ SafePython.html.
33. S. E. Minear. Providing policy control over object operations in a Mach-based
system. In Proceedings of the 5th USENIX Security Symposium, 1995.
34. N. H. Minsky and V. Ungureanu. Unified support for heterogenous security policies
in distributed systems. In Proceedings of the 7th USENIX Security Symposium,
pages 131-142, January 1998.
35. J. K. Ousterhout, J. Y. Levy, and B. B. Welch. The Safe-Tel security model. In
Proceedings of the 23rd USENIX Annual Technical Conference, 1998.
36. R. Rashid, A. Tevanian Jr., M. Young, D. Golub, D. Baron, D. Black, W. J.
Bolosky, and J. Chew. Machine-independent virtual memory management for
paged uniprocessor and multiprocessor architectures. IEEE Transactions on Com-
puters, 37(8):896-908, August 1988.
37. J. H. Saltzer and M. D. Schroeder. The protection of information in computer
systems. Proceedings of the IEEE, 63(9):1278-1308, September 1975.
38. R. Sandhu. Rationale for the RBAC96 family of access control models. In Pro-
ceedings of the 1st Workshop on Role-Based Access Control, 1995.
39. R. Sandhu. Role activation hierarchies. In Proceedings of the Third Workshop on
Role-Based Access Control, 1998.
40. R. S. Sandhu, V. Bhamidipati, E. Coyne, S. Ganta, and C. Youman. The ARBAC97
model for role-based administration of roles: preliminary description and outline.
In Proceedings of the Second Workshop on Role-Based Access Control, pages 41-50,
1997.
41. R. S. Sandhu, E. J. Coyne, H. L. Feinstein, and C. E. Youman. Role-based access
control models. IEEE Computer, 29(2):38-47, February 1996.
42. M. I. Seltzer, Y. Endo, C. Small, and K. A. Smith. Deahng with disaster: Surviving
misbehaved kernel extensions. In Proceedings of the 2nd Conference on Operating
Systems Design and Implementation, pages 213-227, 1996.
43. D. Thomsen, D. O'Brien, and J. Bogle. Role based access control framework for
network enterprises. In Proceedings of the Fourteenth Computer Security Applica-
tions Conference, 1998.
 316

44. J. Vochteloo, K. Elphinstone, S. Russell, and G. Heiser. Protection domain

tensions in Mungi. In Proceedings of the Fifth International Workshop on O
Orientation in Operating Systems, pages 161-165, October 1996.
45. D. S. Wallach and E. W. Felten. Understanding Java stack introspection.
Proceedings of IEEE Symposium on Security and Privacy, 1998.
46. P. Wayner. Agents Unleashed. AP Professional, 1995.
47. M. V. Wilkes and R. M. Needham. The Cambridge CAP Computer and Its C
ating System. North Holland, 1979.
48. W. Wulf, E. Cohen, W. Corwin, A. Jones, R. Levin, C. Pierson, and F. Poll
HYDRA: The kernel of a multiprocessor operating system. Communications o
ACM, 17(6):337-345, June 1974.
49. M. E. Zurko and R. Simon. User-centered security. In Proceedings of the 1996
Security Paradigms Workshop, 1996.
 Providing Policy-Neutral and Transparent
Access Control in Extensible Systems

R o b e r t Grimm and Brian N. Bershad

A b s t r a c t . Extensible systems, such as Java or the SPIN extensible op-

erating system, allow for units of code, or extensions, to be added to a
running system in almost arbitrary fashion. Extensions closely interact
through low-latency, but type-safe interfaces to form a tightly integrated
system. As extensions can come from arbitrary sources, not all of whom
can be trusted to conform to an organization's security policy, such struc-
turing raises the question of how security constraints are enforced in an
extensible system. In this paper, we present an access control mecha-
nism for extensible systems to address this problem. Our access control
mechanism decomposes access control into a policy-neutral enforcement
manager and a security policy manager, and it is transparent to exten-
sions in the absence of security violations. It structures the system into
protection domains, enforces protection domains through access control
checks, and performs auditing of system operations. The access control
mechanism works by inspecting extensions for their types and operations
to determine which abstractions require protection, and by redirecting
procedure or method invocations to inject access control operations into
the system. We describe the design of this access control mechanism,
present an implementation within the SPIN extensible operating sys-
tem, and provide a qualitative as well as quantitative evaluation of the
mechanism.

1 Introduction
Extensible systems, such as Java [18,25] or SPIN [6], promise more power an
flexibility, and thus enable new applications such as smart clients [48] or ai
tive networks [44]. Extensible systems are best characterized by their suppoi
for dynamically composing units of code, called extensions in this paper. ]
these systems, extensions can be added to a running system in almost arbitral
fashion, and they interact through low-latency, but type-safe interfaces with eac
other. Extensions and the core system services are typically co-located within tl
same address space, and form a tightly integrated system. Consequently, extei
sible systems differ fundamentaUy from conventional systems, such as Unix [2£
which rely on processes executing under the control of a privileged kernel.
As a result of this structuring, system security becomes an i m p o r t a n t cha
lenge, and access control becomes a fundamental requirement for the success
extensible systems. As system security is customarily expressed through prote
tion domains [22,38], an access control mechanism must:
 318

- structure the system into protection domains (which are an orthogonal con-
cept to conventional address spaces),
- enforce these domains through access control checks,
- support auditing of system operations.

Furthermore, an access control mechanism must address the fact that extensions
often originate from other networked computers and are untrusted, yet execute
as an integral part of an extensible system and interact closely with other ex-
tensions.
In this paper, we present an access control mechanism for extensible systems
that meets the above requisites. We build on the idea of separating policy and
enforcement first explored by the DTOS effort [30,34,40,39], and introduce a
mechanism that not only separates policy from enforcement, but also access con-
trol from the actual functionality of the system. The access control mechanism
is based on a simple, yet powerful model for the interaction between its policy-
neutral enforcement manager and a given security policy, and is transparent to
extensions and the core system services in the absence of security violations.
It works by inspecting extensions for their types and operations to determine
which abstractions require protection, and by redirecting procedure or method
invocations to inject access control operations into the system.
The access control mechanism provides three types of access control opera-
tions. The operations are (1) explicit protection domain transfers to delineate the
protection domains of an extensible system, (2) access checks to control which
code can be executed and which arguments can be passed between protection
domains, and (3) auditing to provide a trace of system operations. The access
control mechanism works at the granularity of individual procedures (or, ob-
ject methods), and provides precise control over extensions and the core system
services alike.
Access control and its enforcement is but one aspect of the overall security
of an extensible system. Other important issues, such as the specification of
security policies, or the expression and transfer of credentials for extensions
are only touched upon or not discussed at all in this paper. Furthermore, we
assume the existence of some means, such as digital signatures, for authenticating
both extensions and users. These issues are orthogonal to access control, and we
believe that a simple, yet powerful access control mechanism, as presented in
this paper, can serve as a solid foundation for future work on other aspects of
security in extensible systems.
The remainder of this paper is structured as follows: Section 2 elaborates
on the goals of our access control mechanism, and Sect. 3 describes its design.
Section 4 presents the implementation of our access control mechanism within
the SPIN extensible operating system. Section 5 reflects on our experiences with
designing and implementing our access control mechanism, and Sect. 6 presents
a detailed performance analysis of the implementation. Section 7 reviews related
work, and Sect. 8 outlines future directions for our research into the security of
extensible systems. Finally, Sect. 9 concludes this paper.
 319

2 Goals
An access control mechanism for an extensible system must impose additional
structure onto the system. But, at the same time, it should only impose as much
structure as strictly necessary to preserve the advantages of an extensible system.
Based on this realization, we identify four goals which inform the design of our
system.
Separate access control and functionality. The access control mechanism
should separate policy and enforcement from the actual code of the system and
extensions. This separation of access control and functionality supports chang-
ing security policies without requiring access to source code. This is especially
important for large computer networks, such as the Internet, where the same
extension may execute on different systems with different security requirements,
and where source code typically is not available. This goal does not prevent the
programmer who writes an extension from defining (part of) the security policy
for that extension. However, it calls for a separate specification of such policy,
comparable to an interface specification which offers a distinct and concise de-
scription of the abstractions found in a unit of code. This policy specification
may then be loaded into an extensible system as the extension is loaded.
Separate policy and enforcement. The mechanism should separate the security
policy from its actual enforcement. This separation of policy and enforcement
allows for changing security policies without requiring intrinsic changes to the
core services of the extensible system itself. Rather, the security policy is pro-
vided by a (trusted) extension, and, as a result, the access control mechanism
leverages the advantages of an extensible system and becomes extensible itself.
Use a simple, yet expressive model. The mechanism should rely on a simple
model of protection that covers a wide range of possible security policies, includ-
ing policies that change over time or depend on the history of the system. This
goal ensures that the access control mechanism can strictly enforce a wide range
of security policies, and that the security policy has control over all relevant
aspects of access control. At the same time, it favors simplicity over complex
interactions between security policy and its enforcement.
Enforce transparently. The mechanism should be transparent to extensions
and the core system services, in that they should not need to interact with it
as long as no violations of the security policy occur. This goal ensures that
the mechanism actually provides a clean separation of security policy, enforce-
ment, and functionality. Furthermore, it provides support for legacy code (to
a degree), and enables aggressive, policy-specific optimizations that reduce the
performance overhead of access control. At the same time, it guarantees that
extensions are notified of security faults, and can implement their own failure
model. Consequently, this goal attempts to reduce the access control interface,
as seen by extensions, to handling a program fault such as division by zero or
dereferencing a NIL reference.
The above four goals, taken together, call for a design that isolates functional-
ity, security policy, and enforcement in an extensible system, and that provides a
clear specification for their interaction. In other words, the goals call for an access
 320

control mechanism that combines the extension itself, the security constraints
for the extension as specified by the programmer, and a site's security policy
to produce a secure extension. At the same time, the mechanism is not limited
to changing only the extension as a result of this combination process, but can
impose security constraints on other parts of the extensible system as well. This
process of combining functionality and security to provide access control in an
extensible system is illustrated in Fig. 1.

Security
Policy

Security
Constraints Access
Control
Mechanism
Extension

Secure
Extension

Extensible System

Fig. 1. Overview of access control in an extensible system. The access control mecha-
nism combines the extension itself, the security constraints for the extension as specified
by the programmer, and a site's security policy to place a secure version of the extension
into the extensible system.

A design that addresses the four goals effectively defines the protocol by which
the security policy and the access control mechanism interact, and by which, if
necessary, extensions are notified of security-relevant events. As such, this proto-
col is an internal protocol. In other words, the abstractions used for expressing
protection domains and access control checks need not be, and probably should
not be, the same abstractions presented by the security policy. It is the respon-
sibility of a security policy manager to provide users and system administrators
with a high-level and user-friendly view of system security.

2.1 Examples
As long as extensions, such as Java applets in the sandbox model, use only a few,
selected core services, providing protection in an extensible system reduces to
isolating extensions from each other and performing access control checks in the
core services. However, for many real-world applications of extensibility, such
a protection scheme is clearly insufficient as extensions use some parts of the
system and, in turn, are used by other parts. For example, an extension may
 321

provide a new file system implementation, such as a log-structured file system,

offer additional functionality for existing file systems, such as compression or
encryption, or support higher-level abstractions, such as transactions, on top of
the storage services. An extension may also implement new networking proto-
cols, such as multicast, or higher-level communication services, such as a remote
procedure call package or an object request broker (ORB), on top of the existing
networking stack.
Prom a security viewpoint, the programmer who writes such an extension
will want to protect the resources used by that extension. So, for a transaction
manager, she would like to ensure that the files or disk extents used for storing
transaction data can only be accessed through the transaction manager. And, for
an ORB, she would like to ensure that the network port used for communicating
with other nodes can not be accessed by other extensions. A simple way to
implement these security constraints is to place the transaction manager or ORB
into its own protection domain and to use access checks in the storage services or
networking stack to protect the resources they use. The security constraints in
these examples thus not only affect the service provided by the extension itself,
but also cover other services of an extensible system. At the same time, overall
security in an extensible system requires coalescing the constraints for several
extensions. Consequently, separating the specification of security constraints and
functionality would clearly aid in providing security for an extensible system.
In addition to the programmer, the administrator of an extensible system
may want to impose additional restrictions on an extension. For example, she
may want to restrict how other extensions can call on the transaction manager in
order to ensure that only a transaction's initiator can actually commit it. Or, she
may require auditing of the transaction manager's operations to ensure that a log
record is generated if the commit operation is not performed by a transaction's
initiator. Alternatively, the administrator may want to impose a security policy
that conflicts with the security constraints expressed by the programmer. For
example, she may want to integrate the ORB into the same protection domain as
the networking stack, since the ORB is the only means for remote communication
in an installation (such as a corporate intranet), and providing access control on
the ORB is adequate for security. As illustrated by these examples, the security
policy for an extensible system varies according to the requirements of a specific
installation, even if the functionality does not change. It is thus not sufficient to
only separate access control from functionality, but also necessary to separate
the security policy from its enforcement.
So far, we have argued for a clean separation of security policy, functionality,
and enforcement, and for a system that transparently manages security. However,
extensions need to be notified of failures so that they can implement their own
failure model. For example, the transaction manager might decide to abort the
offending transaction, or the ORB may need to clean up the internal state of
the corresponding connection. It is thus important that access control is only
transparent in the absence of failures, and that extensions are notified of security
violations.
 322

3 Design

The design of our access control mechanism divides access control in an exten-
sible system into an enforcement manager and a security policy manager. The
enforcement manager is part of the core services of the extensible system. It
provides information on the types and operations of an extension, and redirects
procedure or method invocations to perform access control operations. The se-
curity policy manager is provided by a trusted extension, and determines the
actual security policy for the system. It decides which procedures require which
access control operations, and performs the actual mediation for access control.
This structure is illustrated in Fig. 2.

1
1
Security Policy
Manager Extensions

Enforcement
Manager

Reflection,
Interposition Core System

Fig. 2. Structure of the access control mechanism. The enforcement manager is part
of the core system services, provides information on the types and operations of an
extension (reflection), and redirects procedure or method invocations (interposition) to
ensure that a given security policy is actually enforced onto the system. The security
policy manager is a trusted extension, determines which abstractions require which
access control operations, and performs the actual mediation.

The protocol that determines the interaction between the enforcement and
the security policy manager relies on two abstractions, namely security identi-
fiers and sets of permissions, or access modes. Security identifiers are associated
with both subjects and objects, and represent privilege and access constraints.
Permissions are associated with operations, and represent the right to perform
an operation. The enforcement manager maintains the association of subjects
and objects with security identifiers, and performs access control checks based
on access modes. But, it does not interpret security identifiers and access modes,
as their meaning is determined by the security policy manager which performs
the actual mediation.
As extensible systems feature a considerably different structuring from tra-
ditional systems such as Unix, it is necessary to define the exact meaning of
subjects and objects. We treat threads in an extensible system as subjects, as
they are the only active entities, and all other entities, including extensions, as
 323

objects. This is not to say that subjects only represent the principal that cre-
ated a thread. Rather, the rights of a subject depend on the current protection
domain, i.e. the extension whose code the thread is currently executing, and,
possibly, on previous protection domains, i.e. the history of extensions whose
code the thread executed before entering the current extension. Furthermore,
while we treat extensions as objects, they are subject to a somewhat different
form of access control than other objects in an extensible system.

3.1 Access Control on Extensions

Conceptually, access control determines whether a subject can legally execute

some operation on some object. Access control on extensions differs from this
concept in that it is sometimes necessary to control how extensions interact with
each other. Specifically, this is the case at link-time: The extension to be loaded
into the system needs to be linked against other extensions, whose interfaces
it will execute and extend. It is thus necessary, at link-time, to provide access
control over which interfaces a given extension can link against for execution and
extending [41,35]. Enforcing this link-time control over extensions is important,
since it presents a first line of defense against unauthorized access (after all, if
an extension can not link against an interface, it can not directly use it), and
since it may result in opportunities for optimizing away dynamic access control
operations.
Link-time control over extensions can, however, be expressed through regular
access control checks by enforcing checks on linkage operations, and by executing
these operations within a protection domain appropriate for the extension to be
linked. To impose access control checks on linkage operations, the enforcement
manager injects the appropriate checks into the linking service during system
start-up. To execute these operations within a protection domain appropriate for
an extension, the loader spawns a new thread for each extension to be loaded into
the system, which then performs the actual linkage operations as well as other
necessary initialization. The initial security identifier of this thread represents
the corresponding protection domain. It is determined by the security policy
manager based on an extension's signature, and is associated with the thread
and the procedures of the extension by the enforcement manager.
When loading an extension, the enforcement manager also determines the
types and operations exported by that extension, and passes this information to
the security poUcy manager. Based on this information and an extension's signa-
ture, the security policy manager determines which operations and types require
which access control operations. The security policy manager, in turn, instructs
the enforcement manager to provide security identifiers for an extension's types,
and to inject access control operations into the extension and other parts of the
system, if necessary. Once the actual linking of an extension is complete and the
appropriate access control operations have been injected into the system, the
extension is fully and securely integrated into the system and its code can now
be executed.
 324

3.2 Access Control Operations

The enforcement manager supports three types of access control operations. The
operations are (1) protection domain transfers to structure the system into pro-
tection domains, (2) access control checks to enforce these protection domains,
and (3) auditing to provide a trace of system operations. Protection domain
transfers change the protection domain associated with a thread, based on the
current protection domain of a thread and on the procedure that is about to
be invoked. Access checks determine whether the current subject is allowed to
execute a procedure at all, and control the passing of arguments and results.
For each argument that is passed into a procedure, and for each result that is
passed back from the procedure, access checks determine whether the subject
has sufficient rights for the corresponding object. Finally, auditing generates a
log-entry for each procedure invocation, and serves as an execution trace of the
system.
When instructing the enforcement manager to perform access control oper-
ations on a given procedure, the security policy manager specifies the types of
access control operations, i.e. any combination of protection domain transfer,
access checks, and auditing. For access checks, it also specifies the required ac-
cess modes, one for the procedure itself, one for each argument, and one for each
result.
The access control operations are ordered as follows. Before a given procedure
is executed, the enforcement manager first performs access checks, then a pro-
tection domain transfer, and, finally, auditing, which also records failed access
checks. On return from the procedure, the enforcement manager first performs
the reverse protection domain transfer, then access checks on the results, and,
finally, auditing, which, again, also records failed access checks.
To perform the access control operations, the enforcement manager requires
three mappings between security identifiers, types, and access modes. These
mappings are used to communicate a site's security policy between security
policy manager and enforcement manager. Using SID for security identifiers.
T Y P E for types as defined by the extensible system, and A C C E S S M O D E for access
modes, the three mappings are:

(1) SID X SID --^ SID

(2) SID X SID --> ACCESSMODE
(3) SID X T Y P E --^ SID

The first mapping is used for protection domain transfers. It maps the current
security identifier of a thread and the security identifier of the procedure that is
about to be called into the new security identifier of the thread. The enforcement
manager associates the thread with the new security identifier before control
passes into the actual procedure, and it restores the original security identifier
upon completion of the procedure.
The second mapping is used for access checks. It maps the security identifier
of a thread and the security identifier of an object into an access mode represent-
ing the max;imum rights the subject has on the object. The enforcement manager
 325

verifies that the maximal access mode contains all permissions of the required
access mode, as specified by the security policy manager when requesting the
access check.

The third mapping is used for the creation of objects. It maps the security
identifier of a thread that is about to create an object and the type of that
object into the security identifier for the newly created object. The enforcement
manager associates newly created objects with the resulting security identifier.
A simplification of this mapping may omit the object type from the mapping,
and simply map all objects created by a thread into the same security identifier.

Both variations of the third mapping provide relatively coarse-grained con-

trol over the security identifiers associated with objects, and are clearly insuf-
ficient for some services. For example, a file server typically executes threads
within its own protection domain but may need to associate different files with
different security identifiers. Thus, to support trusted services that provide finer-
grained control over the security identifiers associated with objects, the enforce-
ment manager provides an interface through which a trusted security service
can change the security identifier of an object. In the example of a file server,
this interface could be used to map files to security identifiers similar to the
name-based security attributes in domain and type enforcement [3,2].

New subjects, that is freshly spawned threads, are associated with the same
security identifier as the spawning thread so that they possess the same privi-
leges. An exception to this rule occurs for threads that are created to link and
initialize extensions (as discussed above), and for threads that are created when
a user logs into the system. In the latter case, an appropriate form of authenti-
cation (such as a password) establishes the identity of the user to the security
policy manager, and the enforcement manager associates the thread with the
corresponding security identifier.

Depending on the complexity of the security policy implemented by the se-

curity policy manager, lookup operations for the three mappings may incur a
relatively high performance overhead. Consequently, the enforcement manager
caches individual entries in the three mappings, which reduces the frequency
with which the security policy manager needs to resolve entries and therefore
the overall performance overhead of access control operations. The security pol-
icy manager has full control over this mediation cache. It sets the overall size of
the cache, can remove any entry from the cache at any time, and also flush the
entire cache. Furthermore, for any lookup operation on any of the mappings, it
specifies whether that particular entry can be cached and, if so, for how long.
For example, for multi-level policies [5,12,7], all mappings can be cached indef-
initely since they never change. For access matrix based policies [22], mappings
generally can be cached. But, if the access matrix is changed, the security policy
manager must remove the corresponding entries from the mediation cache. And,
for policies that depend on the mediation history, mappings may not be cached
at all.
 326

4 Implementation
We have implemented our access control mechanism in the SPIN extensible
operating system [6]. Our access control mechanism does not depend on features
that are unique to SPIN, and could be implemented in other systems. It requires
support for dynamically loading and linking extensions, for multiple concurrent
threads of execution, for determining an extension's types and operations, and
for redirecting procedure or method invocations (for example, by patching object
jump tables either statically or dynamically). Consequently, our access control
mechanism can be implemented in other extensible systems that provide these
features, such as Java.
Our implementation is guided by three constraints. First, it has to correctly
enforce a given security policy as defined by the security policy manager. Second,
it has to be simple and well-structured to allow for validation^ and for easy
transfer to other systems. Third, the implementation should be fast to impose
as little performance overhead as possible.
In SPIN, a statically linked core provides most basic services, including hard-
ware support, the Modula-3 runtime [42,21], the linker/loader [41], threads, and
the event dispatcher [35]. All other services, including networking and file system
support, are provided by dynamically linked extensions. We have implemented
the basic abstractions of our access control mechanism, such as security identi-
fiers and access modes, as well as the enforcement manager as part of this static
core.
Services in the static core are trusted in that, if they misbehave, the security
of the system can be undermined, and the system may even crash. At the same
time, the static core must be protected against dynamically linked extensions
which usually are not trusted. Consequently, the enforcement manager imposes
access control on the core services, including the linker/loader as described in
Sect. 3.1, to protect itself and other core services, and to ensure that only a
trusted extension can define the security policy. User-space applications in SPIN
need not be written in Modula-3, are not guaranteed to be type-safe, and thus
are generally untrusted. They can not access any kernel-level objects directly,
but only through a narrowly defined system call interface, which automatically
subjects them to our access control mechanism.
The implementation of our access control mechanism consists of 1000 lines
of well-documented Modula-3 interfaces and 2400 lines of Modula-3 code, with
an additional 50 lines of changes to other parts of the static SPIN core. The im-
plementation uses the Modula-3 runtime to determine the types and operations
of an extension, and the event dispatcher [35] to inject access control operations
into the system. It defines the abstractions for security identifiers and access
modes. Security identifiers are simply integers. Access modes are immutable ob-
jects, and are represented by a set of simple, pre-defined permissions in addition
to a list of permission objects. The simple permissions provide 64 permissions at
We have not validated the implementation. However, a critical chaxacteristic for any
security mechanism is that it be small and well-structured [38].
 327

a low overhead. The hst of permission objects lets the security policy manager
define additional permissions (where each permission object can represent sev-
eral permissions) by subtyping from an abstract base class, at some performance
cost.

The functionality of the enforcement manager is visible through two separate

interfaces. One interface lets extensions discover the state of system security in
the presence of security faults. The other interface, together with the interface
to the security policy manager, defines the trusted protocol between the secu-
rity policy and the enforcement manager. The interface to the security policy
manager is also defined as part of the static core, but an implementation of
this interface must be provided by a trusted extension outside the static core.
The enforcement manager operates as described in Sect. 3. It uses the simpli-
fied third mapping for assigning objects to security identifiers (see Sect. 3.2),
and thus provides a default security identifier for all objects within a protection
domain.

On object creation, the standard Modula-3 allocator, through a call-back into

the enforcement manager, stores this default security identifier in the header of
the newly allocated object. At the same time, only some types in an extensible
system require access control. For example, an auxiliary object that is only used
within an extension and never passed outside will never require access control.
Thus, to limit the memory overhead of allocating an additional word in each
object header, the security manager can dynamically activate and deactivate
object security for each Modula^3 type individually. Access checks on objects
that are not associated with a security identifier simply fail.

Storing an object's security identifier in the object header considerably sim-

plifies the mapping from objects to security identifiers as the enforcement man-
ager does not need to maintain a separate mapping. For example, when an
unused object is freed by the garbage collector, the corresponding mapping is
deleted with the object and no additional operation needs to be performed by
the enforcement manager. Furthermore, as security identifiers are stored in the
same location as the object itself, the performance overhead to access an object's
security identifier is minimized.

To maintain a thread's security identifier and the corresponding default ob-

ject security identifier, the enforcement manager associates each thread with
a security identifier stack. Each record on this stack contains the two security
identifiers for the subject and its objects. On a protection domain transfer, the
enforcement manager pushes a new record onto the stack before the thread enters
the corresponding procedure or method, and pops the record off the stack when
the thread returns from the procedure or method. Records are pre-allocated in a
global pool to avoid dynamic memory allocation overhead, and are pushed and
popped using atomic enqueue and dequeue operations to avoid the overhead of
locking the global pool.
 328

5 Discussion
By using our access control mechanism, fine-grained security constraints can be
imposed onto an extensible system. However, the expressiveness of our mecha-
nism is limited in that it can not supplant prudent interface design. In particular,
three issues arise, namely the use of abstract data types, the granularity of in-
terfaces, and the effect of calling conventions.
Our access control mechanism provides protection on objects in that it pro-
vides control over which operations a subject can legally execute on an object,
including control over which objects can be passed to and returned from an oper-
ation. To do so, it relies on abstract data types to hide the implementation of an
object. In other words, if the type of an object does not hide its implementation,
it is possible to directly access and modify an object without explicitly invoking
any of the corresponding operations and thus without incurring access control.
The structure of an interface also influences the degree of control attainable
over the operations on an object. In particular, the granularity of an interface,
i.e. how an interface decomposes into individual operations on a type, deter-
mines the granularity of access control. So, an interface with only one operation,
which, like i o c t l in Unix, might use an integer argument to name the actual
operation, allows for much less fine-grained control than an interface with several
independent operations.
The calling convention used for passing arguments to a procedure or object
method affects whether argument passing can be fully controlled. Notably, call-
by-reference grants both caller and callee access to the same variable. As caller
and callee may be in different protection domains, call-by-reference effectively
creates (type-safe) shared memory. In a multi-threaded system, information can
be passed through shared memory at any time, not just on procedure invocation
and return. Consequently, caller and callee need to trust each other on the use
of this shared memory, and access checks on call-by-reference arguments are
not very meaningful. In SPIN, call-by-reference is almost always used to return
additional results from a procedure, as Modula-3 only supports one result value.
This unnecessary use of shared memory could clearly be avoided by supporting
multiple results or thread-safe calling conventions such as call-by-value/result at
the programming language level.
The three issues just discussed are directly related to our access control mech-
anism relying on an extension's interface, that is on the externally visible types
and operations of an extension, to impose access constraints. A more powerful
model could be used to express finer-grained security constraints. And, more
aggressive techniques, such as binary rewriting [45,43,19,37], could be used to
enforce these constraints in an extensible system. But such a system would also
require a considerably more complex design and implementation. At the same
time, an extension's interface is a "natural" basis for access control, as it provides
a concise and well-understood specification of what an extension exports to other
extensions and how it interacts with them. Consequently, we believe that our
access control mechanism strikes a reasonable balance between expressiveness
and complexity.
 329

As our access control mechanism relies on extensions' interfaces to provide

protection for an extensible system, it also requires some means to ensure that
these interfaces are, in fact, respected by the actual code. SPIN uses a type-
safe programming language (Modula-3), and a trusted compiler to provide this
guarantee. As a result, the compiler becomes part of the trusted computing base.
Clearly, it is preferable to establish this guarantee in the extensible system that
actually executes the code, especially for large computer networks. Considerable
work has been devoted to this issue, and viable alternatives include typed byte-
codes [25], proof-carrying code [33], as well as typed assembly language [31]. All
of these efforts are complementary to our own.

6 Performance Evaluation
To determine the performance overhead of our implementation, we evaluate a set
of micro-benchmarks that measure the performance of access control operations.
We also present end-to-end performance results for a web server benchmark. We
collected our measurements on a DEC Alpha AXP 133 MHz 3000/400 worksta-
tion, which is rated at 74 SPECint 92. The machine has 64 MByte of memory,
a 512 KByte unified external cache, and an HP C2247-300 1 GByte disk-drive.
In summary, the micro-benchmarks show that access control operations incur
some latency on trivial operations, while the end-to-end experiment shows that
the overall overhead of access control is in the noise.

6.1 Micro-Benchmarks

To evaluate the performance overhead of access control operations in our access

control mechanism, we execute seven micro-benchmarks. All seven benchmarks
measure the total time for a null procedure call (a procedure that returns im-
mediately and does not perform any work), with and without access control
operations. The first benchmark simply performs a null procedure call with no
arguments. The other six benchmarks additionally perform a protection domain
transfer, an access check on the procedure, and access checks on one, two, four
and eight arguments, respectively.
The performance of the security policy manager is determined by a given
security policy and its implementation. Consequently, for the micro-benchmarks,
we fix the necessary entries in the mediation cache of the enforcement manager
(see Sect. 3.2). The benchmarks thus measure common-case performance, where
the security policy manager is not consulted because the necessary information
is already available within the enforcement manager. Furthermore, benchmarks
that perform access control checks use simple permissions instead of permission
objects (see Sect. 4).
Table 1 shows the performance results for the seven micro-benchmarks. All
numbers are in microseconds and the average of 1000 trials. To determine hot
microprocessor cache performance, we execute one trial to pre-warm the proces-
sor's cache, and then execute it 1000 times in a tight loop, measuring the time
 330

Table 1. Performance numbers for access control operations. All numbers are the mean
of 1000 trials in microseconds. Hot represents hot microprocessor cache performance
and Cold cold microprocessor cache performance.

Hot Cold
Null procedure call 0.1 0.5
Protection domain transfer 4.4 7.8
Access check on procedure 2.8 6.4
Access check on 1 argument 4.0 9.7
Access check on 2 arguments 6.7 12.0
Access check on 4 arguments 12.1 17.7
Access check on 8 arguments 24.0 29.5

at the beginning and at the end of the loop. To determine cold microprocessor
cache performance, we measure the time before and after each trial separately,
and flush both the instruction and data cache on each iteration.
Table 2 shows the instruction breakdown of the common path for protec-
tion domain transfers, excluding the overhead for the event dispatcher (which
amounts to 31 or 48 instructions, depending on the optimizations used within
the event dispatcher [35]). On a protection domain transfer, the enforcement
manager establishes the new protection domain before control passes into the
actual procedure, and restores the original protection domain upon completion
of the procedure. Before entering the procedure, the enforcement manager first
determines the security identifiers of the thread and of the procedure. Then,
based on these security identifiers, it looks up the security identifiers for the
thread and new objects created by the thread in the mediation cache, which re-
quires obtaining a lock for the cache. Next, it sets up a new exception frame, so
that the original protection domain can be restored on an exceptional procedure
exit. Finally, it pushes a new record containing the security identifiers for the
thread and new objects onto the thread's security identifier stack. After leaving
the procedure, the enforcement manager pops the top from the thread's security
identifier stack and removes the exception frame.
Additional experiments show that performing a protection domain transfer
in addition to access checks adds 3.9 microseconds to hot cache performance
and 5.6 microseconds to cold cache performance for those of the above bench-
marks that perform access checks. Furthermore, using permission objects instead
of simple permissions for access checks, where the required permission object
matches the tenth object in the list of legal permission objects (which represents
a pessimistic scenario as each permission object can stand for dozens of indi-
vidual permissions), adds 6.8 microseconds for hot cache performance and 7.0
microseconds for cold cache performance per argument.
The performance results show that access control operations have noticeable
overhead. They thus back our basic premise that access control for extensible
systems should only impose as much structure as strictly necessary. Furthermore,
 331

Table 2. Instruction breakdown of the common path for protection domain transfers
excluding the cost for the event dispatcher. "Overhead" is the overhead of performing
both protection domain changes within their own procedure. The other operations are
explained in the text.

Operation # Instr.

Enter new protection domain

Get thread's security ID 3
Get procedure's security ID 1
Lookup in mediation cache 52
Locking overhead 62
Set up exception frame 7
Push security ID record 26
Overhead 10
Total number of instructions 161

Restore old protection domain

Pop security ID record 22
Remove exception frame 4
Overhead 4
Total number of instructions 30

they underline the need for a design that enables dynamic optimizations which
avoid access control operations whenever possible.

6.2 End-to-End Performance

To evaluate the overall impact of access control on system performance, we

present end-to-end results for a web server benchmark. The web server used for
our experiments is implemented as an in-kernel extension. It uses an NFS client
to read files from our group's file server, and locally caches the file data in a
dedicated cache, backed by a simple, fast extent-based file system. As spawning
new threads in SPIN incurs very little overhead, the web server forks a new
thread for each incoming request. The thread first checks whether the requested
file is available in the local cache, and, if so, sends the file data directly from
the cache. Otherwise, it issues an NFS read request, stores the file in the local
cache, and then sends the data.
Our security policy places the web server into its own protection domain. It
performs access control checks on all NFS and local cache operations. Files in the
local cache are automatically associated with a security identifier as described
in Sect. 3.2. Files in NFS are associated with a security identifier by using a
mapping from the file system name-space to security identifiers (similar to the
one described in [3,2]) to provide fine-grained control over which files are as-
sociated with which security identifier. Since the security policy imposes access
 332

control checks on both the NFS client and the local cache, and since the indi-
vidual threads (spawned to serve requests) can only communicate through NFS
and the local cache, the policy ensures that only authorized files are accessible
through the web server. Furthermore, it makes it possible to securely change
privileges on a per-request basis, either based on a remote login, or based on the
machine from which the request originated.
Our performance benchmark sends http requests from one machine that is
running the benchmark script to another that is running the web server. It reads
the entire SPIN web tree, to a total of 79 files or 5035 KByte of data. We run the
benchmark without access control, as a baseline, as well as with access control,
to measure the end-to-end overhead of our access control mechanism. For each
measurement, we first perform 15 runs of the benchmark to pre-warm the local
cache, and then measure the latency for 20 runs. The average latency for one run
of the benchmark both without and with access control is 16.9 seconds (including
5.4 seconds idle time on the machine running the web server), and the difference
between the two is in the noise. Trials with access control incur a total of 1573
access checks, on average 20 for each file.
The end-to-end performance experiments show that the overhead of access
control operations is negligible for a web server workload. We extrapolate from
this result that other applications will see a very small overhead for other real-
world applications. To better quantify this overhead, we plan to conduct further
experiments in the future that use more complex security policies and require
finer-grained access control operations.

7 Related Work

A considerable body of literature focuses on system protection [22,38] and ap-

propriate security policies. Starting from multi-level security [5,12,7], which has
become part of the U.S. Department of Defense's standard for trusted computer
systems [13], much attention has been directed towards mapping non-military
policies onto multi-level security [26,24], defining alternative policies more suit-
able for commercial applications [8,10,9,3,2,1], and expanding multi-level se-
curity to be more flexible and powerful [27,32].
Based on the realization that no single security policy is appropriate for all
environments, the DTOS effort [30,34,40,39] has the goal of providing a policy-
neutral access control mechanism. As our mechanism builds on this work, we
share with DTOS the same structuring of access control into a security policy
manager and a policy-neutral enforcement manager as well as the same basic
abstractions (security identifiers and permissions). However, as DTOS has been
implemented on top of the Mach micro-kernel, it differs form our mechanism in
that it relies on address-spaces for protection domains, resulting in a relatively
high overhead for changing protection domains. Furthermore, as the DTOS effort
does not separate security from functionality, it uses explicit access checks on
pre-defined permissions for enforcing protection domains, making it impossible
to change or remove access checks.
 333

As reported in [40], adding explicit access checks to the micro-kernel pre-

sented a considerable challenge as it fixed part of the security policy within the
system. Furthermore, as noted in [39], their choice of checking whether a subject
can perform an operation on an object, where the object is the primary argu-
ment to an operation, does not provide sufficient fiexibility, since the security
decision may depend on other parameters to the operation as well. Our access
control mechanism avoids these limitations, as access control operations are dy-
namically specified and injected into the system, and as they are strictly more
expressive.
Due to Java's [18,25] popularity for providing executable content on the
Internet, and prompted by a string of security breaches [11,28] in early versions
of the system, research into protection for extensible systems has mostly focused
on Java. In departure from the original sandbox model, which grants trusted
code full access to the underlying system and untrusted code almost no access,
the Java security architecture is currently being extended [15,16] to allow for
multiple protection domains, provide fine-grained access control primitives, and
support cryptographic protocols.
The basic technique for performing dynamic access checks in Java, called
extended stack introspection, is described in [46]. With this technique, each
extension is implicitly associated with a protection domain, and access checks
essentially take the intersection of all protection domains represented on the
current call-stack to determine if an operation is legal. While extended stack
introspection is sufficiently expressive to provide fine-grained access control [47],
it is closely tied to the stack-based execution model of Java. Furthermore, it
relies on explicit access checks, and thus fails to separate functionality from
protection. Finally, as access checks need to walk the entire call-stack, they can
incur considerable performance overhead [17].
Hagimont and Ismail [20] describe an alternative design for access control in
Java which provides for a separate description of security constraints through an
extended interface definition language. In their design, security constraints are
expressed as part of the interface specification for each extension, and result in
the creation of proxy objects which provide only limited functionality to their
clients. The design essentially provides a form of type hiding [46] at the granu-
larity of entire methods, as the visibility of object methods is controlled by the
security constraints.
In its ability to provide access control at the granularity of object methods,
Hagimont and Ismail's design is similar to CACL [36] which presents a general
protection model for objects. At the same time, CACL offers a more complete
model (which includes explicit representations of the owner of an object and its
implementor) and a more efficient implementation (through object jump tables
instead of proxy objects). The idea of using hmited effective types to avoid
repeated dynamic access checks that determine whether a subject can call on
an object is complimentary to our design. We thus believe that it could be used
to provide an efficient implementation of the enforcement manager in a pure
object-oriented system.
 334

8 F u t u r e Work
The access control mechanism described in this paper provides us with an ideal
test-bed for future research on the security of extensible systems. Specifically,
the policy-neutral and transparent enforcement manager, with its ability to ar-
bitrarily inject protection domains and access checks into an extensible system,
offers us considerable power and flexibility. We are particularly interested in
three areas for future work: First, programmers and security administrators need
to be able to specify security constraints for the code they write and use. We
thus plan to investigate appropriate specification languages that are both user-
friendly (i.e., present a high-level of abstraction) and sufficiently powerful to
conveniently express detailed security policies (i.e., provide enough flexibility).
Second, as extensions often execute in networked environments, a protocol for
the secure expression and transfer of credentials is required. We thus intend to
examine distributed authentication and authorization protocols, such as those
described in [23,4,14], in the context of extensible systems. Finally, as illus-
trated by the micro-benchmarks in Sect. 6, the access control operations show a
relatively high overhead when compared to a simple procedure invocation. We
thus plan to explore aggressive optimizations that avoid dynamic access control
operations whenever possible.

9 Conclusions
The access control mechanism for extensible systems described in this paper
breaks up access control into a policy-neutral enforcement manager and a secu-
rity policy manager, and is transparent to extensions in the absence of security
violations. It structures the system into protection domains through protec-
tion domain transfers, enforces these protection domains through access control
checks, and provides a trace of system operations through auditing. It works by
inspecting extensions for their types and operations to determine what abstrac-
tions require protection, and by redirecting procedure or method invocations to
inject access control operations into the system. The access control mechanism
is based on a simple, yet powerful protocol by which the security policy and the
enforcement manager interact, and by which, if necessary, extensions are notified
of security-relevant events.
The implementation of our access control mechanism within the SPIN ex-
tensible operating system is simple, and, even though the latency of individual
access control operations can be noticeable, shows good end-to-end performance.
Based on our results, we predict that most systems will see a very small overhead
for access control, and thus consider our access control mechanism an eflFective
solution for access control in extensible systems.

Acknowledgments
The research presented in this paper was sponsored by the Defense Advanced
Research Projects Agency, the National Science Foundation and by an equipment
 335

grant from Digital Equipment Corporation. Grimm was partially supported by

fellowships from the Microsoft Corporation and the IBM Corporation. Bershad
was partially supported by a National Science Foundation Presidential Faculty
Fellowship a n d an Office of Naval Research Young Investigator Award.
We t h a n k Dennis Hollingworth and Timothy Redmond at Trusted Informa-
tion Systems for their comments on our design. Marc Fiuczynski, Tian Lim,
Yasushi Saito, Emin Giin Sirer and especially Przemyslaw P a r d y a k at t h e Uni-
versity of Washington were most helpful with various implementation issues
and t h e integration of our access control mechanism into SPIN. Stefan Savage,
Stephen Smalley, Ray Spencer, Amin Vahdat, and t h e anonymous reviewers for
this volume provided valuable feedback on earlier versions of this paper.

References
1. L. Badger, K. A. Oostendorp, W. G. Morrison, K. M. Walker, C. D. Vance, D. L.
Sherman, and D. F. Sterne. DTE Firewalls—Initial Measurement and Evaluation
Report. Technical Report 0632R, Trusted Information Systems, March 1997.
2. L. Badger, D. F. Sterne, D. L. Sherman, K. M. Walker, and S. A. Haghighat.
A Domain and Type Enforcement UNIX Prototype. In Proceedings of the Fifth
USENIX UNIX Security Symposium, pages 127-140, Salt Lake City, Utah, June
1995.
3. L. Badger, D. F. Sterne, D. L. Sherman, K. M. Walker, and S. A. Haghighat.
Practical Domain and Type Enforcement for UNIX. In Proceedings of the 1995
IEEE Symposium on Security and Privacy, pages 66-77, Oakland, California, May
1995.
4. E. Belani, A. Vahdat, T. Anderson, and M. Dahlin. The CRISIS Wide Area
Security Architecture. In Proceedings of the 7th USENIX Security Symposium,
San Antonio, Texas, January 1998.
5. D. E. Bell and L. J. La Padula. Secure Computer System: Unified Exposition
and Multics Interpretation. Technical Report MTR-2997 Rev. 1, The MITRE
Corporation, Bedford, Massachusetts, March 1976. Also ADA023588, National
Technical Information Service.
6. B. N. Bershad, S. Savage, P. Pardyak, E. G. Sirer, M. Fiuczynski, D. Becker,
S. Eggers, and C. Chambers. Extensibility, Safety and Performance in the SPIN
Operating System. In Proceedings of the 15th Symposium on Operating Systems
Principles, pages 267-284, Copper Mountain, Colorado, December 1995.
7. K. J. Biba. Integrity Considerations for Secure Computer Systems. Technical Re-
port MTR-3153 Rev. 1, The MITRE Corporation, Bedford, Massachusetts, April
1977. Also ADA039324, National Technical Information Service.
8. W. E. Boebert and R. Y. Kain. A Practical Alternative to Hierarchical Integrity
Policies. In Proceedings of the 17th National Computer Security Conference, pages
18-27, Gaithersburg, Maryland, 1985.
9. D. F. C. Brewer and M. J. Nash. The Chinese Wall Security Policy. In Proceedings
of the 1989 IEEE Symposium on Security and Privacy, pages 206-214, Oakland,
California, May 1989.
10. D. D. Clark and D. R. Wilson. A Comparison of Commercial and Military Com-
puter Security Policies. In Proceedings of the 1987 IEEE Symposium on Security
and Privacy, pages 184-194, Oakland, California, April 1987.
 336

11. D. Dean, E. W. Felten, and D. S. Wallach. Java Security: Prom HotJava to

Netscape and Beyond. In Proceedings of the 1996 IEEE Symposium on Security
and Privacy, pages 190-200, Oakland, California, May 1996.
12. D. E. Denning. A Lattice Model of Secure Information Flow. Communications of
the ACM, 19(5):236-243, May 1976.
13. Department of Defense Computer Security Center. Department of Defense Trusted
Computer System Evaluation Criteria, December 1985. Department of Defense
Standard DoD 5200.28-STD.
14. C. M. Ellison, B. Prantz, B. Lampson, R. Rivest, B. M. Thomas, and T. Ylo-
nen. SPKI Certificate Theory. Technical Report draft-ietf-spki-cert-theory-04.txt,
Internet Engineering Task Porce, November 1998.
15. L. Gong. Java Security: Present and Near Future. IEEE Micro, 17(3):14-19,
May/June 1997.
16. L. Gong, M. Mueller, H. Prafullchandra, and R. Schemers. Going Beyond the
Sandbox: An Overview of the New Security Architecture in the Java Development
Kit 1.2. In Proceedings of the USENIX Symposium on Internet Technologies and
Systems, pages 103-112, Monterey, Cahfornia, December 1997.
17. L. Gong and R. Schemers. Implementing Protection Domains the Java Develop-
ment Kit 1.2. In Proceedings of the Internet Society Symposium on Network and
Distributed System Security, San Diego, California, March 1998.
18. J. Gosling, B. Joy, and G. Steele. The Java Language Specification. Addison-
Wesley, Reading, Massachusetts, 1996.
19. S. L. Graham, S. Lucco, and R. Wahbe. Adaptable Binary Programs. In Pro-
ceedings of the 1995 USENIX Technical Conference, pages 315-325, New Orleans,
Louisiana, January 1995.
20. D. Hagimont and L. Ismail. A Protection Scheme for Mobile Agents on Java. In
Proceedings of the Third Annual ACM/IEEE International Conference on Mobile
Computing and Networking, Budapest, Hungary, September 1997.
21. W. C. Hsieh, M. E. Piuczynski, C. Garrett, S. Savage, D. Becker, and B. N. Ber-
shad. Language Support for Extensible Operating Systems. In Proceedings of
the Workshop on Compiler Support for System Software, pages 127-133, Tucson,
Arizona, February 1996.
22. B. W. Lampson. Protection. In Proceedings of the Fifth Princeton Symposium on
Information Sciences and Systems, pages 437-443, Princeton, New Jersey, March
1971. Reprinted in Operating Systems Review, 8(l):18-24, January 1974.
23. B. W. Lampson, M. Abadi, M. Burrows, and E. Wobber. Authentication in Dis-
tributed Systems: Theory and Practice. ACM Transactions on Computer Systems,
10(4):265-310, November 1992.
24. T. M. P. Lee. Using Mandatory Integrity to Enforce "Commercial" Security. In
Proceedings of the 1988 IEEE Symposium on Security and Privacy, pages 140-146,
Oakland, California, April 1988.
25. T. Lindholm and F. Yelhn. The Java Virtual Machine Specification. Addison-
Wesley, Reading, Massachusetts, 1996.
26. S. B. Lipner. Non-Discretionary Controls for Commercial Applications. In Pro-
ceedings of the 1982 Symposium on Security and Privacy, pages 2-10, Oakland,
California, April 1982.
27. C. J. McCollum, J. R. Messing, and L. Notargiacomo. Beyond the Pale of MAC
and DAC—Defining New Forms of Access Control. In Proceedings of the 1990
IEEE Symposium on Research in Security and Privacy, pages 190-200, Oakland,
California, May 1990.
 337

28. G. McGraw and E. W. Felten. Java Security: Hostile Applets, Holes and Antidotes.
Wiley Computer Publishing, John Wiley &; Sons, Inc., New York, New York, 1997.
29. M. K. McKusick, K. Bostic, M. J. Kaxels, and J. S. Quarterman. The Design
and Implementation of the 4-4BSD Operating System. Addison-Wesley Publishing
Company, Reading, Massachusetts, 1996.
30. S. E. Minear. Providing Policy Control Over Object Operations in a Mach Based
System. In Proceedings of the Fifth USENIX UNIX Security Symposium, pages
141-156, Salt Lake City, Utah, June 1995.
31. G. Morrisett, D. Walker, K. Crary, and N. Glew. Prom System F to Typed Assem-
bly Language. In Proceedings of the 25th Symposium on Principles of Programming
Languages, San Diego, California, January 1998.
32. A. C. Myers and B. Liskov. A Decentralized Model for Information Flow Control.
In Proceedings of the 16th Symposium on Operating Systems Principles, pages 129-
142, Saint-Malo, Prance, October 1997.
33. G. C. Necula and P. Lee. Safe Kernel Extensions Without Run-Time Checking.
In Proceedings of the Second Symposium on Operating Systems Design and Imple-
mentation, pages 229-243, Seattle, Washington, October 1996.
34. D, Olawsky, T. Fine, E. Schneider, and R. Spencer. Developing and Using a "Policy
Neutral" Access Control Policy. In Proceedings of the New Security Paradigms
Workshop, September 1996.
35. P. Pardyak and B. N. Bershad. Dynamic Binding for an Extensible System. In
Proceedings of the Second Symposium on Operating Systems Design and Imple-
mentation, pages 201-212, Seattle, Washington, October 1996.
36. J. Richardson, P. Schwarz, and L.-F. Cabrera. CACL: Efficient Fine-Grained Pro-
tection for Objects. In Proceedings of the Conference on Object-Oriented Pro-
gramming Systems, Languages, and Applications '92, pages 263-275, Vancouver,
Canada, October 1992.
37. T. Romer, G. Voelker, D. Lee, A. Woman, W. Wong, H. Levy, B. N. Bershad, and
B. Chen. Instrumentation and Optimization of Win32/Intel Executables Using
Etch. In Proceedings of the USENIX Windows NT Workshop, pages 1-8, Seattle,
Washington, August 1997.
38. J. H. Saltzer and M. D. Schroeder. The Protection of Information in Computer
Systems. Proceedings of the IEEE, 63(9):1278-1308, September 1975.
39. Secure Computing Corporation. DTOS General System Security and Assurability
Assessment Report. Technical Report DTOS CDRL AOll, Secure Computing
Corporation, Secure Computing Corporation, 2675 Long Lake Road, Roseville,
Minnesota 55113-2536, June 1997.
40. Secure Computing Corporation. DTOS Lessons Learned Report. Technical Report
DTOS CDRL A008, Secure Computing Corporation, Secure Computing Corpora-
tion, 2675 Long Lake Road, Roseville, Minnesota 55113-2536, June 1997.
41. E. G. Sirer, M. Fiuczynski, P. Pardyak, and B. N. Bershad. Safe Dynamic Linking
in an Extensible Operating System. In Proceedings of the Workshop on Compiler
Support for System Software, pages 134-140, Tucson, Arizona, February 1996.
42. E. G. Sirer, S. Savage, P. Pardyak, G. P. DeFouw, M. A. Alapat, and B. N. Bershad.
Writing an Operating System with Modula-3. In Proceedings of the Workshop on
Compiler Support for System Software, pages 141-148, Tucson, Arizona, February
1996.
43. A. Srivastava and A. Eustace. ATOM: A System for Building Customized Program
Analysis Tools. In Proceedings of the ACM SIGPLAN '94 Conference on Program-
ming Language Design and Implementation, pages 196-205, Orlando, Florida, June
1994.
 338

44. D. L. Tennenhouse, J. M. Smith, W. D. Sincoskie, D. J. Wetherall, and G. J.

Minden. A Survey of Active Network Research. IEEE Communications Magazine,
25(l):80-86, January 1997.
45. R. Wahbe, S. Lucco, T. E. Anderson, and S. L. Graham. Efficient Software-Based
Fault Isolation. In Proceedings of the 14th Symposium on Operating Systems Prin-
ciples, pages 203-216, Ashville, North Carolina, December 1993.
46. D. S. Wallach, D. Balfanz, D. Dean, and E. W. Felten. Extensible Security Ar-
chitectures for Java. In Proceedings of the 16th Symposium on Operating Systems
Principles, pages 116-128, Saint-Malo, Prance, October 1997.
47. D. S. Wallach and E. W. Felten. Understanding Java Stack Inspection. In Proceed-
ings of the 1998 IEEE Symposium on Security and Privacy, pages 52-63, Oakland,
California, May 1998.
48. C. Yoshikawa, B. Chun, P. Eastham, A. Vahdat, T. Anderson, and D. Culler. Using
Smart Clients to Build Scalable Services. In Proceedings of the 1997 USENIX
Technical Conference, pages 105-117, Anaheim, California, January 1997.
 Interposition Agents:
Transparently Interposing User Code
at the System Interface*

Michael B. Jones

Abstract. Many contemporary operating systems utilize a system call

interface between the operating system and its clients. Increasing num-
bers of systems are providing low-level mechanisms for intercepting and
handling system calls in user code. Nonetheless, they typically provide
no higher-level tools or abstractions for effectively utilizing these mecha-
nisms. Using them has typically required reimplementation of a substan-
tial portion of the system interface from scratch, making the use of such
facilities unwieldy at best.
This paper presents a toolkit that substantially increases the ease of in-
terposing user code between clients and instances of the system interface
by allowing such code to be written in terms of the high-level objects
provided by this interface, rather than in terms of the intercepted sys-
tem calls themselves. This toolkit helps enable new interposition agents
to be written, many of which would not otherwise have been attempted.
This toolkit has also been used to construct several agents including;
system call tracing tools, file reference tracing tools, and customizable
filesystem views. Examples of other agents that could be built include:
protected environments for running untrusted binaries, logical devices
implemented entirely in user space, transparent data compression and/or
encryption agents, transactional software environments, and emulators
for other operating system environments.

1 Introduction
1.1 Terminology

Many contemporary operating systems provide an interface between user code

and the operating system services based on special "system calls". One can view
the system interface as simply a special form of structured communication chan-
nel on which messages are sent, allowing such operations as interposing programs
t h a t record or modify the communications t h a t take place on this channel. In
this paper, such a program t h a t b o t h uses and provides the system interface will
be referred t o as a "system interface interposition agent" or simply as an "agent"
for short.

This paper was originally printed in Proceedings of the 14th ACM Symposium on
Operating Systems Principles, pages 80-93. Asheville, NC, December, 1993 ©1993
ACM
 340

1.2 Overview
This paper presents a toolkit that substantially increases the ease of interposing
user code between clients and instances of the system interface by allowing such
code to be written in terms of the high-level objects provided by this interface,
rather than in terms of the intercepted system calls themselves. Providing an
object-oriented toolkit exposing the multiple layers of abstraction present in
the system interface provides a useful set of tools and interfaces at each level.
Different agents can thus exploit the toolkit objects best suited to their individual
needs. Consequently, substantial amounts of toolkit code are able to be reused
when constructing different agents. Furthermore, having such a toolkit enables
new system interface implementations to be written, many of which would not
otherwise have been attempted.
Just as interposition is successfully used today to extend operating system
interfaces based on such communication-based facilities as pipes, sockets, and
inter-process communication channels, interposition can also be successfully used
to extend the system interface. In this way, the known benefits of interposition
can also be extended to the domain of the system interface.

1.3 Examples
The following figures should help clarify both the system interface and interpo-
sition.

mail make

9WK^ssomiy*^^iaoxismxm,\
Operating System Kernel

Operating System Interface

f open () , r e a d O , s t a t O , fork () ,
k i l l ( ) , _exit{), signals, . . . )

Fig. 1. Kernel provides instances of system interface

Figure 1 depicts uses of the system interface without interposition. In this

view, the kernel^ provides all instances of the operating system interface.
The term "kernel" is used throughout this paper to refer to the default or lowest-level
implementation of the operating system in question. While this implementation is
often run in processor kernel space, this need not be the case, as in the Mach 3.0
Unix Server/Emulator [16].
 341

Application Program

WyAOrAmim^m«/AyjrMgmiK\

Your code here!

WSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSi

Operating System Kernel

Operating System Interface

Fig. 2. User code interposed at system interface

Figure 2 depicts the ability to transparently interpose user code that both
uses and implements the operating system interface between an unmodified ap-
plication program and the operating system kernel.

Untrusted
emacs Binary

^Aimium'.
Compress/ Restricted
csh Uncompress Environment

weimxGiKimi^ryyyyyjyryjxm V^;j?jj???JWA
Operating System Kernel

Fig. 3. Kernel and agents provide instances of system interface

Figure 3 depicts uses of the system interface with interposition. Here, both
the kernel and interposition agents provide instances of the operating system
interface.
 342

HP-UX
Compiler
v/xf/xofxaa^^
HP-UX
make Emulator

'W/////////A V^>^WWM>y
mail Agent Implementing
Customized Filesystem View

W/Z/MM V///////?/////A *m!msme6xa>mKL

Operating System Kernel

Fig. 4. Agents can share state and provide multiple instances of system interface

Figure 4 depicts more uses of the system interface with interposition. In this
view agents, like the kernel, can share state and provide multiple instances of
the operating system interface.

1.4 Motivation
Today, agents are regularly written to be interposed on simple communication-
based interfaces such as pipes and sockets. Similarly, the toolkit makes it possible
to easily write agents to be interposed on the system interface.
Interposition can be used to provide programming facilities that would oth-
erwise not be available. In particular, it can allow for a multiplicity of simul-
taneously coexisting implementations of the system call services, which in turn
may utilize one another without requiring changes to existing client binaries and
without modifying the underlying kernel to support each implementation.
Alternate system call implementations can be used to provide a number of
services not typically available on system call-based operating systems. Some
examples include:
— System Call Tracing and Monitoring Facilities: Debuggers and pro-
gram trace facilities can be constructed that allow monitoring of a program's
use of system services in a easily customizable manner.
— Emulation of Other Operating Systems: Alternate system call imple-
mentations can be used to concurrently run binaries from variant operating
systems on the same platform. For instance, it could be used to run UL-
TRIX [13], HP-UX [10], or UNIX System V [3] binaries in a Mach/BSD
environment.
— Protected Environments for Running Untrusted Binaries: A wrap-
per environment can be constructed that allows untrusted, possibly mali-
cious, binaries to be run within a restricted environment that monitors and
 343

emulates the actions they take, possibly without actually performing them,
and limits the resources they can use in such a way that the untrusted bina-
ries are unaware of the restrictions. A wide variety of monitoring and emu-
lating schemes are possible from simple automatic resource restriction envi-
ronments to heuristic evaluations of the target program's behavior, possibly
including interactive decisions made by human beings during the protected
execution. This is particularly timely in today's environments of increased
software sharing with the potential for viruses and Trojan horses.
— Transactional Software Environments: Applications can be constructed
that provide an environment in which changes to persistent state made by
unmodified programs can be emulated and performed transactionally. For
instance, a simple "run transaction" command could be constructed that
runs arbitrary unmodified programs (e.g., / b i n / c s h ) such that all persistent
execution side effects (e.g., filesystem writes) are remembered and appear
within the transactional environment to have been performed normally, but
where in actuality the user is presented with a "commit" or "abort" choice at
the end of such a session. Indeed, one such transactional program invocation
could occur within another, transparently providing nested transactions.
- Alternate or Enhanced Semantics: Environments can be constructed
that provide alternate or enhanced semantics for unmodified binaries. One
such enhancement in which people have expressed interest is the ability to
"mount" a search list of directories in the filesystem name space such that
the union of their contents appears to reside in a single directory. This could
be used in a software development environment to allow distinct source and
object directories to appear as a single directory when running make.

1.5 Problems with Existing S y s t e m s

Increasing numbers of operating systems are providing low-level mechanisms
for intercepting system calls. Having these low-level mechanisms makes writing
interposition agents possible. For instance, Mach [1,16] provides the intercep-
tion facilities used for this work, SunOS version 4 [44] provides new p t r a c e O
operations used by the t r a c e utility, and UNIX System V.4 [4] provides new
/ p r o c operations used by the t r u s s utility. Nonetheless, they typically provide
no higher-level tools or abstractions for effectively utilizing these mechanisms,
making the use of such facilities unwieldy at best.
Part of the difficulty with writing system call interposition agents in the past
has been that no one set of interfaces is appropriate across a range of such agents
other than the lowest level system call interception services. Different agents in-
teract with different subsets of the operating system interface in widely different
ways to do different things. Building an agent often requires implementation
of a substantial portion of the system interface. Yet, only the bare minimum
interception facilities have been available, providing only the lowest common de-
nominator that is minimally necessary. Consequently, each agent has typically
been constructed completely from scratch. No leverage was gained from the work
done on other agents.
 344

1.6 Key Insight

The key insight that enabled me to gain leverage on the problem of writing sys-
tem interface interposition agents for the 4.3BSD [25] interface is as follows: while
the 4.3BSD system interface contains a large number of different system calls,
it contains a relatively small number of abstractions ('whose behavior is largely
independent). (In 4.3BSD, the primary system interface abstractions are path-
names, descriptors, processes, process groups, files, directories, symbolic links,
pipes, sockets, signals, devices, users, groups, permissions, and time.) Further-
more, most calls manipulate only a few of these abstractions.
Thus, it should be possible to construct a toolkit that presents these ab-
stractions as objects in an object-oriented programming language. Such a toolkit
would then be able to support the substantial commonalities present in different
agents through code reuse, while also supporting the diversity of different kinds
of agents through inheritance.

2 Research Overview
2.1 Design Goals

The four main goals of the toolkit were:

1. Unmodified System: Unmodified apphcations should be able to be run

under agents. Similarly, the underlying kernel should not require changes to
support each different agent (although the kernel may have to be modified
once in order to provide support for system call interception, etc. so that
agents can be written at all).
2. Completeness: Agents should be able to both use and provide the entire
system interface. This includes not only the set of requests from applications
to the system (i.e., the system calls) but also the set of upcalls that the
system can make upon the applications (i.e., the signals).
3. Appropriate Code Size: The amount of new code necessary to implement
an agent using the toolkit should only be proportional to the new functional-
ity to be implemented by the agent — not to the size of the system interface.
The toolkit should provide whatever boilerplate and tools are necessary to
write agents at levels of abstraction that are appropriate for the agent func-
tionality, rather than having to write each agent at the raw system call level.
4. Performance: The performance impact of running an application under an
agent should be negligible.

2.2 Design and Structure of the Toolkit

I have designed and built a toolkit on top of the Mach 2.5 system call interception
mechanism [1,5,16] that can be used to interpose user code on the 4.3BSD [25]
system call interface. The toolkit currently runs on the Intel 386/486 and the
VAX. The toolkit is implemented in C + + with small amounts of C and assembly
 345

language as necessary. Multi-threaded hybrid 4.3BSD/Mach 2.5 programs are

not currently supported.
As a consequence of using the Mach 2.5 system call interception mechanism,
which redirects system calls to handler routines in the same address space, in-
terposition agents reside in the same address spaces as their client processes.
The lowest layers of the toolkit hides this Mach-specific choice, allowing agents
to be constructed that could be located either in the same or different addresses
spaces as their clients.
This toolkit is structured in an object-oriented manner, allowing agents to
be written in terms of several different layers of objects by utilizing inheritance.
Abstractions exposed at different toolkit layers currently include the filesystem
name space, pathnames, directories, file descriptors and the associated descrip-
tor name space, open objects referenced by descriptors, and signals, as well as
the system calls themselves. (These abstractions are discussed further in Section
2.3.) Support for additional abstractions can be incrementally added as needed
by writing new toolkit objects that represent the new abstractions and by using
derived versions of the existing toolkit objects that reference the new abstrac-
tions through the new objects. Indeed, the current toolkit was constructed via
exactly this kind of stepwise refinement, with useful toolkit objects being pro-
duced at each step. The structure of the toolkit permits agents to be written
in terms of whatever system interface abstractions are appropriate to the tasks
they perform. Just as derived objects are used to introduce new toolkit function-
ality, interposition agents change the behavior of particular system abstractions
by using agent-specific derived versions of the toolkit objects representing those
abstractions.
Different interposition agents need to affect different components of the sys-
tem call interface in substantially different ways and at different levels of abstrac-
tion. For instance, a system call monitoring/profiling agent needs to manipulate
the system calls themselves, whereas an agent providing alternate user filesystem
views needs to manipulate higher-level objects such as pathnames and possibly
file descriptors. The agent writer decides what layers of toolkit objects are ap-
propriate to the particular task and includes only those toolkit objects. Default
implementations of the included objects provide the normal behavior of the ab-
stractions they represent. This allows derived agent-specific versions of toolkit
objects to inherit this behavior, while adding new behavior in the implementa-
tions of the derived objects. I believe that the failure to provide such multi-layer
interfaces by past system call interception mechanisms has made them less useful
than they might otherwise have been.

2.3 Toolkit Layers

Figure 5 presents a diagram of the primary classes currently provided with the
interposition toolkit. Indented classes are subclasses of the classes above. Arrows
indicate the use of one class by another. Many of these classes are explained in
more detail in this section.
 346

numeric_syscall
bsd_numeric_syscall

i
symbolic_syscall
_desc_symbolic_syscall
path_symbolic_syscall

descriptor_set pathname_s et
open_descriptor_set

descriptor pathname
open_descriptor

open_object
directory

Fig. 5. Primary interposition toolkit classes

The lowest layers of the toolkit perform such functions as agent invocation,
system call interception, incoming signal handling, performing system calls on
behalf of the agent, and delivering signals to applications running under agent
code. Unlike the higher levels of the toolkit, these layers are sometimes highly
operating system specific and also contain machine specific code.
These layers hide the mechanisms used to intercept system calls and signals,
those that are used to call down from an agent to the next level system interface,
and those that are used to send a signal from an agent up to the application
program. These layers also hide such details as whether the agent resides in the
same address space as the application program or whether it resides in a separate
address space. These layers are referred to as the boilerplate layers. These layers
are not normally used directly by interposition agents.
The lowest (or zeroth) layer of the toolkit which is directly used by any in-
terposition agents presents the system interface as a single entry point accepting
vectors of untyped numeric arguments. It provides the ability to register for spe-
cific numeric system calls to be intercepted and for incoming signal handlers to
be registered. This layer is referred to as the numeric system call layer.
Example interfaces provided by the numeric system call layer are as follows:
 347

class immeric_syscall •[
public:
virtual int syscall(int number, int args[],
int rv[2], void *regs);
virtual void init(char *agentargv[]);
virtual void signal_handler(int sig, int code,
struct sigcontext *context);
void register_interest(int number);
void register_interest_range(int low, int high);
};

For instance, using just the numeric system call layer, by using a derived
version of the numeric_syscall class with an agent-specific s y s c a l l O method,
an agent writer could trivially write an agent that printed the arguments of a
chosen set of system calls as uninterpreted numeric values. As another example,
one range of system call numbers could be remapped to calls on a different range
at this level.
The first layer of the toolkit intended for direct use by most interposition
agents presents the system interface as a set of system call methods on a system
interface object. When this layer is used by an agent, application system calls are
mapped into invocations on the system call methods of this object. (This map-
ping is itself done by a toolkit-supplied derived version of the numeric_syscall
object.) This layer is referred to as the symbolic system call layer.
Example interfaces provided by the symbolic system call layer are as follows:

c l a s s symbolic_syscall {
public:
v i r t u a l void i n i t ( c h a r * a g e n t a r g v [ ] ) ;
v i r t u a l void i n i t _ c h i l d ( ) ;

virtual int sys_exit(int status, int rv[2]);

v i r t u a l int sys_fork(int r v [ 2 ] ) ;
v i r t u a l i n t s y s _ r e a d ( i n t fd, void *buf, i n t c n t , i n t r v [ 2 ] ) ;
. . . entries for all other ^.3BSD system calls ...

v i r t u a l i n t unknown_syscall(int number, i n t * a r g s , i n t r v [ 2 ] ,
s t r u c t emul_regs * r e g s ) ;
v i r t u a l void s i g n a l _ h a n d l e r ( i n t s i g , i n t code,
s t r u c t sigcontext * c o n t e x t ) ;
};

Agents can interpose on individual system calls by using a derived version

of the symbolic_syscall object with agent-specific methods corresponding to
the system calls to be intercepted. For instance, the timex agent, which is de-
scribed in Section 3.3, changes the apparent time of day by using a derived
symbolic_syscall object with a new gettimeofdayO method. Likewise, the
 348

t r a c e agent, described in Section 3.3, prints the arguments to each executed sys-
tem call in a human-readable from individual system call methods in a derived
syinbolic_syscall object.
The second layer of the toolkit is structured around the primary abstrac-
tions provided by the system call interface. In 4.3BSD, these include path-
names, file descriptors, processes, and process groups. This layer presents the
system interface as sets of methods on objects representing these abstractions.
Toolkit objects currently provided at this level are the filesystem name space
(pathncune_set), resolved pathnames (pathname), the file descriptor name space
( d e s c r i p t o r _ s e t ) , active file descriptors (descriptor), and reference counted
open objects (open_object). Such operations as filesystem name space trans-
formations and filesystem usage monitoring are done at this level.
For example, agents can interpose on pathname operations by using derived
versions of two classes: pathname_set and pathname. The pathname_set class
provides operations that affect the set of pathnames, i.e., those that create or
remove pathnames. The pathname class provides operations on the objects ref-
erenced by the pathnames.
Example interfaces provided by the pathname_set class are as follows:
c l a s s pathncime_set : p u b l i c d e s c r i p t o r _ s e t {
protected:
v i r t u a l i n t getpn(char *path, i n t f l a g s , pathname **pn);
public:
v i r t u a l void i n i t ( c h a r * a g e n t a r g v [ ] ,
c l a s s PATH_SYMBOLIC_BASE *path_sym);

/ / System c a l l s with knowledge of pathnames

v i r t u a l i n t open(char *path, i n t f l a g s , i n t mode, i n t r v [ 2 ] ) ;
v i r t u a l i n t l i n k ( c h a r *path, char *newpath, i n t r v [ 2 ] ) ;
v i r t u a l i n t unlink(char *path, i n t r v [ 2 ] ) ;
. . . entries for other ^.3BSD system calls using pathnames ...
};
Example interfaces provided by the pathname class are as follows:
c l a s s pathncime {
public:
v i r t u a l i n t open(int f l a g s , i n t mode, i n t r v [ 2 ] ,
OPEN_OBJECT_CLASS **oo);
v i r t u a l i n t link(pathname *newpn, i n t r v [ 2 ] ) ;
virtual int unlink(int rv[2]);
. . . entries for other 4-3BSD system calls referencing objects
via pathnames ...
};
The key to both of these interrelated classes is the g e t p n O operation, which
looks up a pathname string and resolves it to a reference to a pathname ob-
ject. The default implementation of all the pathname_set system call methods
 349

simply resolves their pathname strings to pathname objects using getpnO and
then invokes the corresponding pathname method on the resulting object. The
pathname method is responsible for actually performing the requested operation
on the object referenced by the pathname.
With the getpnO operation to encapsulate pathname lookup, it is possible
for agents to supply derived versions of the pathname_set object with a new
getpnO implementation that modifies the treatment of all pathnames. For in-
stance, this can be used to logically rearrange the pathname space, as was done
by the union agent (described in Section 3.3). Likewise, it provides a central
point for name reference data collection, as was done by the df s_trace agent
(described in Section 3.5).
A third set of toolkit layers focuses on secondary objects provided by the
system call interface, which are normally accessed via primary objects. Such
objects include files, directories, symbohc links, devices, pipes, and sockets. These
layers present the system interface as sets of methods on objects, with specialized
operations for particular classes of objects. The only toolkit object currently
provided at this level is the open directory d i r e c t o r y object. Operations that
are specific to these secondary objects such as directory content transformations
are done at this level.
For example, agents can interpose on directory operations by using derived
versions of the d i r e c t o r y class. The directory class is itself a derived version of
the open_object class (one of the second layer classes for file descriptor oper-
ations), since directory operations are a special case of operations that can be
performed on file descriptors.
Example interfaces provided by the d i r e c t o r y class are as follows:

c l a s s d i r e c t o r y : public OPEN_OBJECT_CLASS {
public:
virtual int next_direntry();
s t r u c t d i r e c t * d i r e n t r y ; / / Set by n e x t _ d i r e n t r y ( )

public:
virtual int read(void *buf, int cnt, int rv[2]);
virtual int lseek(off_t offset, int whence, int rv[2]);
virtual int getdirentries(void *buf, int cnt, long *basep,
int rv[2]);
};

Just as the getpnO method encapsulated pathname resolution, the

n e x t _ d i r e n t r y O method encapsulates the iteration of individual directory en-
tries implicit in reading the contents of a directory. This allows the union
agent (described in Section 3.3) to make it appear that the full contents of
a set of directories is actually present in a single directory by providing a
new next _ d i r e n t r y 0 function that iterates over the contents of each mem-
ber directory. (And yes, that iteration itself is accomplished via the underlying
n e x t _ d i r e n t r y implementations.)
 350

2.4 Using the Toolkit to Build Agents

As I built the toolkit, I also used it to implement several interposition agents.

These agents provide:

— System Call and Resource Usage Monitoring: This demonstrates the

ability to intercept the full system call interface.
— User Configurable Filesystem Views: This demonstrates the ability to
transparently assign new interpretations to filesystem pathnames.
— File Reference Tracing Tools: that are compatible with existing tools [30]
originally implemented for use by the Coda [38,23] filesystem project: this
provides a basis for comparing a best available equivalent implementation to
a facility provided by an agent.

3 Results

3.1 Goal: Unmodified System

Unmodified Applications

Agents constructed using the system interface interposition toolkit can load
and run unmodified 4.3BSD binaries. No recompilation or relinking is neces-
sary. Thus, agents can be used for all program binaries — not just those for
which sources or object files are available.
Applications do not have to be adapted to or modified for particular agents.
Indeed, the presence of agents should be transparent to applications.'^

Unmodified Kernel

Agents constructed using the system interface interposition toolkit do not require
any agent-specific kernel modifications. Instead, they use general system call
handling facilities that are provided by the kernel in order to implement all
agent-specific system call behavior. Also, a general agent loader program is used
to invoke arbitrary agents, which are compiled separately from the agent loader.
The Mach 2.5 kernel used for this work contains a primitive,
t a s k _ s e t _ e m u l a t i o n ( ) , that allows 4.3BSD system calls to be redirected for
execution in user space. Another primitive, h t g _ u n i x _ s y s c a l l ( ) , permits calls
to be made on the underlying 4.3BSD system call implementation even though
those calls are being redirected.

^ Of course, an application that is intent on determining if it is running under an

agent probably can, if only by probing memory or performing precise performance
measurements.
 351

3.2 Goal: Completeness

Agents constructed using the system interface interposition toolkit can both use
and provide the entire 4.3BSD system interface. This includes not only the sys-
tem calls, but also the signals. Thus, both the downward path (from applications
to agents and from agents to the underlying system implementation) and the
upward path (from the underlying implementation to agents and from agents to
applications) are fully supported.
Completeness gives two desirable results:
1. All programs can potentially be run under agents. By contrast, if complete-
ness did not hold, there would have been two classes of programs: those that
used a restricted set of features that agents could handle, and those that
used features that agents could not handle. The interposition toolkit avoids
these problems.
2. Agents can potentially modify all aspects of the system interface. Agents are
not restricted to modifying only subsets of the system behavior. For instance,
it would have been easy to envision similar systems in which agents could
modify the behavior of system calls, but not incoming signals.

3.3 Goal: Appropriate Code Size

Table 1 lists the source code sizes of three different agents, broken down into
statements of toolkit code used, and statements of agent specific code.^ These
agents were chosen to provide a cross section of different interposition agents,
ranging from the very simple to the fairly complex and using different portions
of the interposition toolkit. Each of these agents is discussed in turn.

Sizes of Agents
Agent Toolkit Agent Total
Name Statements Statements Statements
timex 2467 35 2502
trace 2467 1348 3815
union 3977 166 4143

Table 1. Sizes of agents, measured in semicolons

Size of the Timex Agent

The timex agent changes the apparent time of day. It is built upon the symbolic
system call and lower levels of the toolkit (see Section 2.3). The toolkit code used
^ Note: The actual metric used was to count semicolons. For C and C++, this gives a
better measure of the actual number of statements present in the code than counting
lines in the source files.
 352

for this agent contains 2467 statements. The code specific to this agent consists of
only two routines: a new derived implementation of the gettimeofdayO system
call and an initialization routine to accept the desired effective time of day from
the command line. This code contains only 35 statements.
The core of the timex agent is as follows:

c l a s s timex_symbolic_syscall : p u b l i c symbolic_syscall {
public:
v i r t u a l void i n i t ( c h a r * a g e n t a r g v [ ] ) ;
v i r t u a l i n t sys_gettimeofday(struct timeval * t p ,
s t r u c t timezone * t z p , i n t r v [ 2 ] ) ;
private:
int offset; / / Difference between r e a l eind funky time
};

int timex_symbolic_syscall::sys_gettimeofday(
struct timeval *tp, struct timezone *tzp, int rv[2])
{
int ret;
ret = symbolic_syscall::sys_gettimeofday(tp, tzp, rv);
if (ret >= 0 && tp) {
tp->tv_sec += offset;
}
return ret;
}
The new code necessary to construct the timex agent using the toolkit con-
sists only of the implementation of the new functionality. Inheritance from toolkit
objects is used to obtain implementations of all system interface behaviors that
remain unchanged.

Size of the Trace Agent

The t r a c e agent traces the execution of client processes, printing each system
call made and signal received. Like the timex agent, it is built upon the symbolic
system call and lower levels of the toolkit, which contain 2467 statements. How-
ever, the code specific to this agent is much larger, containing 1348 statements.
The reason for this is simple: unhke the timex agent, the new work of the t r a c e
agent is proportional to the size of the entire system interface. Derived versions
of each of the 114 4.3BSD system calls plus the signal handler are needed to
print each call name and arguments, since each call has a different name and
typically takes different parameters. Even so, the new code contains less than 12
statements per
system call, 10 of which typically are of the form:

v i r t u a l i n t s y s _ r e a d ( i n t fd, void +buf, i n t c n t , i n t r v [ 2 ] ) ;

(line from TRACE_SYMBOLIC_CLASS c l a s s declaration)
 353

i n t TRACE_SYMBOLIC_CLASS::sys_read(int fd, void *buf, i n t c n t ,

int rv[2])
{
register int ret;
print_start();
fprintfCf, "read(7.d, Ox'/.x, Ox7.x) ... ]
n", fd, buf, cnt);
fflush(f);
ret = TRACE_SYMBOLIC_BASE::sys_read(fd, buf, cnt, rv);
print_start();
f p r i n t f C f , " . . . readC/.d, Ox'/.x, Ox'/.x) ->", fd, buf, cnt);
print_retx(ret, rv);
return r e t ;
}
As with the timex agent, the new code necessary to construct the t r a c e agent
using the toolkit consists only of the implementation of the new functionality.
Inheritance from toolkit objects is used to obtain implementations of all system
interface behaviors that remain unchanged.

Size of the Union Agent

The union agent implements union directories, which provide the ability to
view the contents of lists of actual directories as if their contents were merged
into single "union" directories. It is built using toolkit objects for pathnames,
directories, and descriptors, as well as the symbolic system call and lower levels
of the toolkit. The toolkit code used for this agent contains 3977 statements.
The code specific to this agent consists of three things: a derived form of the
toolkit pathname object that maps operations using names of union directories
to operations on the underlying objects, a derived form of the toolkit directory
object that makes it possible to list the logical contents of a union directory via
g e t d i r e n t r i e s O and related calls, and an initiahzation routine that accepts
specifications of the desired union directories from the command line. Yet, this
new code contains only 166 statements.
The new code necessary to construct the union agent using the toolkit con-
sists only of the implementation of the new functionality. As with the other
agents, inheritance from toolkit objects is used to obtain implementations of all
system interface behaviors that remain unchanged.

Size Results
The above examples demonstrate several results pertaining the code size of
agents written using the interposition toolkit. One result is that the size of
the toolkit code dominates the size of agent code for simple agents. Using the
toolkit, the amount of new code to perform useful modifications of the system
interface semantics can be small.
 354

Furthermore, the amount of agent specific code can be proportional to the

new functionahty being implemented by the agent, rather than proportional
to the number of system calls affected. For instance, even though the union
directory agent needs to change the behavior of all 30 calls that use pathnames,
and all 48 calls that use descriptors, or 70 calls in all (eight of which use both),
it is written in terms of toolkit objects that encapsulate the behavior of these
abstractions, rather than in terms of the system calls that use them. Thus, the
agent specific code need only implement the new functionality since the toolkit
provides sufficient underpinnings to make this possible.
Finally, there can be substantial code reuse between different agents. All
the agents listed above were able to use the symbolic system call and lower
levels of the toolkit, consisting of 2467 statements. Both the union agent and
df s_trace agenf* are also able to use the descriptor, open object, and pathname
levels of the toolkit, reusing a total of 3977 statements. Rather than modifying
an implementation of the system interface in order to augment its behavior,
the toolkit makes it possible to implement derived versions of the base toolkit
objects, allowing the base toolkit objects that implement the system interface
to be reused.

3.4 Goal: Performance

Application Performance Data

This section presents the performance of running two applications under sev-
eral different agents. The two applications chosen differ both in their system
call usage and their structure: One makes moderate use of system calls and is
structured as a single process; the other makes heavy use of system calls and is
structured as a collection of related processes. Likewise, the agents chosen range
from very simple to fairly complex. The results are discussed in Section 3.4.

Performance of Formatting a Document

Table 2 presents the elapsed time that it takes to format a preliminary draft
of my dissertation with Scribe [36] on a VAX 6250 both using no agent and
when run under three different agents. In each case, the time presented is the
average of nine successive runs done after an initial run from which the time was
discarded.
This task requires 716 system calls. When run without any agents, it takes
131.5 seconds of elapsed time.
When run under the simplest agent, timex, an additional half second of
overhead is added, giving an effective additional cost of under one half percent
of the base run time. When run under t r a c e , an extra 3.5 seconds of overhead
are introduced. Furthermore, when run under union, the most complex agent

The df s.trace agent implements file reference tracing tools that aie compatible
with existing tools [30] originally implemented for use by the Coda [38, 23] filesystem
project. This agent is discussed further in Section 3.5.
 355

Format my dissertation
agent Name Seconds Slowdown
None 131.5 —
timex 132.0 0.5%
trace 135.0 2.5%
union 136.5 3.5%

Table 2. Time to format my dissertation

considered, there is only an additional 5.0 seconds, giving an effective agent cost
of 3.5% of the base run time.
It comes as no surprise that t r a c e , while conceptually simple, incurs percep-
tible overheads. Each system call made by the application to the t r a c e agent
results in at least an additional two w r i t e () system calls in order to write the
trace output.^

Performance of Compiling C Programs

Table 3 presents the elapsed time that it takes to compile eight small C programs
using Make [15] and the GNU C compiler [40] on a 25MHz Intel 486. In each
case, the time presented is the average of nine successive runs done after an
initial run from which the time was discarded.
To do this, Make runs the GNU C compiler, which in turn runs the C pre-
processor, the C code generator, the assembler, and the linker for each program.
This task requires a total of 11877 system calls, including 64 f o r k ( ) / e x e c v e ( )
pairs. When run without any agents, it takes 16.0 seconds of elapsed time.

Make 8 programs
Agent Name Seconds Slowdown
None 16.0 ~
timex 19.0 19%
trace 33.0 107%
union 29.0 82%

Table 3. Time to make 8 programs

When run under the simplest agent, timex, an additional three seconds of
overhead are added, giving an effective additional cost of 19% of the base run-
time. When run under union, which interposes on most of the system calls and
which uses several additional layers of toolkit abstractions, the additional over-
head beyond the no agent case is 13.0 seconds, giving an effective additional cost

^ Trace output is not buffered across system calls so it will not be lost if the process
is killed.
 356

of 82% of the base runtime. When run under t r a c e , an additional 17.0 seconds
of run time are incurred, yielding a slowdown of 107%.
Again, it comes as no surprise that union introduces more overhead than
timex. It interposes on the vast majority of the system calls, unlike timex,
which interposes on only the bare minimum plus gettimeofdayO. Also, union
uses several additional layers of implementation abstractions not used by timex.
As with the previous application, the larger slowdown for t r a c e is unsur-
prising. Given the large number of system calls made by this application and
the additional two w r i t e ( ) operations performed per application system call for
writing the trace log, the log output time constitutes a significant portion of the
slowdown.
An analysis of low-level performance characteristics is presented in Sections
3.5 and 3.5.

Application Performance Results

The application performance data demonstrates that the performance impact
of running an application under an agent is very agent and application specific.
The performance impact of the example agents upon formatting my dissertation
was practically neghgible, ranging from 0.5% for the timex agent to 2.5% for
the t r a c e agent. However, the performance impact of the example agents upon
making the eight small C programs was significant, ranging 19% for timex to
107% for t r a c e . Unsurprisingly, different programs place different demands upon
the system interface, and different agents add different overheads.
The good news is that the additional overhead of using an agent can be small
relative to the time spent by applications doing actual work. Even though no
performance tuning has been done on the current toolkit implementation, the
overheads already appear to be acceptable for certain classes of applications and
agents.
Furthermore, the agent overheads are of a pay-per-use nature. Calls not in-
tercepted by interposition agents go directly to the underlying system and result
in no additional overhead.
Finally, even though some performance impact is clearly inevitable, pre-
sumedly the agent will have been used because it provides some benefit. For
instance, agents may provide features not otherwise available, or they may pro-
vide a more cost-effective means of implementing a desired set of features than
is otherwise available. The performance "lost" by using an interposition agent
can bring other types of gains.®

3.5 Other Results

Low-level Performance Measurements
Micro Performance Data
This section presents the performance of several low-level operations used to
For a discussion on the tradeoffs of using interposition agents, see Section 5.3.
 357

implement interposition and of several commonly used system calls both without
and with interposition.
Table 4 presents the performance of several low-level operations used to im-
plement interposition. All measurements were taken on a 25MHz Intel 486 run-
ning Mach 2.5 version X144. The code measured was compiled with gcc or g-l-+
version 1.37 with debugging (-g) symbols present.

Performance of Low Level operations

Operation usee
C procedure call with 1 axg, result 1.22
C-I--I- virtual procedure call with 1 arg, result 1.94
Intercept and return from system call 30
htg_unix_syscall() overhead 37

Table 4. Performance measurements of individual low-level operations

Table 5 presents the performance of several commonly used system calls

both without interposition and when a simple interposition agent is used. The
interposition agent, time_symbolic, intercepts each system call, decodes each
call and arguments, and calls C-I--I- virtual procedures corresponding to each
system call. These procedures just take the default action for each system call;
they make the same system call on the next level of the system (the instance of
the system interface on which the agent is being run). This allows the minimum
toolkit overhead for each intercepted system call to be easily measured. Measured
pathnames are in a UFS [27] filesystem and contain 6 pathname components.

Performance of Syst em Calls

Operation usee fisec fisec
without agent with agent toolkit overhead
getpidO 25 170 145
gettimeofdayO 47 214 167
fstatO 54 220 166
readO IK of data 370 579 209
statO 892 1101 209
forkO, waitO, _exit() 10350 22350 12000
execveO 9720 20000 10280

Table 5. Performance measurements of individual system calls

Micro Performance Results

Two times from Table 4 are particularly significant. First, it takes 30/isecto
 358

intercept a system call, save the register state, call a system call dispatching
routine, return from the dispatching routine, load a new register state, and return
from the intercepted system call. This provides a lower bound on the total cost
of any system call implemented by an interposition agent.
Second, using h t g _ u n i x _ s y s c a l l ( ) ^ to make a system call adds 37/isec6f
overhead beyond the normal cost of the system call. This provides a lower bound
on the additional cost for an agent to make a system call that otherwise would
be intercepted by the agent.
Thus, any system call intercepted by an agent that then makes the same
system call as part of the intercepted system call's implementation will take at
least 67/iseclonger than the same system call would have if made with no agent
present. Comparing the 67/isec6verhead to the normal costs of some commonly
used system calls (found in Table 5) helps puts this cost in perspective.
The 67/isec6verhead is quite significant when compared to the execution
times of simple calls such as g e t p i d O or gettimeof day ( ) , which take 25/isecand
47/isec; respectively, without an agent. It becomes less so when compared to
r e a d O or s t a t ( ) , which take 370/isecand 892/isecto execute in the cases mea-
sured without an agent.
Hence, the impact will always be significant on small calls that do very little
work; it can at least potentially be insignificant for calls that do real work.
In practice, of course, the overheads of actual interposition agents are higher
than the 67^sectheoretical minimum. The actual overheads for most system calls
implemented using the symbolic system call toolkit level (see Section 2.3) range
from about 140 to 210/isec; as per Table 5. Overheads for f o r k O and execveO
are significantly greater, adding approximately 10 milliseconds to both, roughly
doubling their costs.
The execveO call is more expensive than most because it must be completely
reimplemented by the toolkit from lower-level primitives, unlike most calls where
the version provided by the underlying implementation can be used. The under-
lying implementation's execveO call can not be used because it clears its caller's
address space. While the application must be reloaded, the agent needs to be
preserved. Thus, the extra expense of execveO is due to having to individually
perform such operations as clearing the caller's address space, closing a subset of
the descriptors, resetting signal handlers, reading the program file, loading the
executable image into the address space, loading the arguments onto the stack,
setting the registers, and transferring control into the loaded image, all of which
are normally done by a single execveO call. Likewise, f o r k O and _ e x i t ( )
are more expensive due to the number of additional bookkeeping operations
required.
While the current overheads certainly leave room for optimization (starting
with compiling the agents with optimization on), they are already low enough
to be unimportant for many applications and agents, as discussed in Section 3.4.

^ The htg_nnix_syscall () facihty permits calls to be made on the underlying 4.3BSD

system call implementation even though those calls are being intercepted.
 359

Finally, it should be stressed that these performance numbers are highly

dependent upon the specific interposition mechanism used. In particular, they
are strongly shaped by agents residing in the address spaces of their clients.

Portability
The interposition toolkit should port to similar systems such as SunOS and
UNIX System V. Despite toolkit dependencies on such Mach facilities as the
particular system call interception mechanism used, all such dependencies were
carefully encapsulated within the lowest (boilerplate) layers of the toolkit. None
of the toolkit layers above the boilerplate directly depends on Mach-specific ser-
vices. Higher toolkit layers, while being intentionally 4.3BSD specific, contain
no such dependencies. This 4.3BSD dependency imposes at most minor porta-
bility concerns to other UNIX-derived systems, given their common lineage and
resulting substantial similarity. Thus, it should be possible to port the toolkit by
replacing the Mach-dependent portions of the boilerplate layers with equivalent
services provided by the target environment.
Likewise, interposition agents written for the toolkit should also readily port.
Even if there are differences between the system interfaces, the toolkit port
should be able to shield the agents from these differences, unless, of course, the
agents are directly using the facilities which differ.
One caveat, however, is probably in order. While a port of the toolkit could
shield interposition agents from low-level system interface differences, it certainly
can not shield them from system performance differences. If the
toolkit is ported to systems that provide significantly slower system call in-
terception mechanisms (as, for instance, mechanisms based on UNIX signals are
likely to be), then some agents which previously exhibited acceptable slowdown
might exhibit unacceptable slowdown when ported.

Comparison to a Best Available Implementation

As an element of this research, an interposition agent (df s_trace) was con-
structed that implements file reference tracing tools that are compatible with
the existing kernel-based DFSTrace [30] tracing tools originally implemented for
use by the Coda [38,23] filesystem project. This was done to provide a realistic
basis for comparing a best available implementation of a task that was imple-
mented without benefit of the toolkit with an equivalent interposition agent
constructed using the toolkit.
While this comparison is not presented in detail here* several of the resulting
conclusions are worth noting. Two key points made evident by the comparison
are:

- Agents can be easy to construct. It appears that constructing an interposition

agent that provides an enhanced implementation of the system interface can

See [22] for a detailed presentation of this compaxison.

 360

be at least as easy and possibly easier than modifying an existing operating

system implementation to perform the equivalent functions.
- Agents may not perform as well as monolithic implementations. Agents that
need to access resources maintained by the underlying operating system im-
plementation will be limited in their performance by the overhead involved
in crossing the system interface boundary in order to access those resources.
Hence, the best monolithic implementation of a given facility needing access
to system resources will always perform better than the best interposition-
based implementation of the same facility. For instance, the kernel-based
DFSTrace tools in the default mode caused a 3.0% slowdown while exe-
cuting the AFS filesystem performance benchmarks [19]. The agent-based
implementation caused a 64% slowdown under the same workload.

Other points also made evident by the comparison are:

— Agents can be as small as the equivalent changes to a monolithic implemen-

tation. Interposition agents built using the interposition toolkit can contain
no more new code than the amount of code changed or added to a monolithic
system implementation to implement equivalent facilities. For instance, the
original DFSTrace kernel and user data collection code contains 1627 state-
ments, compared to 1584 statements for the agent-based implementation.
- Agents can be better structured than monolithic implementations. Interpo-
sition agents built using the interposition toolkit can be more logically struc-
tured and be more portable than a monolithic implementation of equivalent
facilities.
— Agents require no system modifications. Unlike monolithic implementations,
where providing an enhanced implementation of a system often requires mod-
ifying the code implementing the system, interposition agents can provide
enhanced implementations as an independent layer requiring no modifica-
tions to the underlying system. For instance, the original DFSTrace imple-
mentation required the modification of 26 kernel files in order to insert data
collection code under conditional compilation switches; the agent-based im-
plementation required no modifications to existing code since inheritance was
used to add functionality. Also, the kernel-based implementation uses four
machine-dependent files per machine type; the agent-based implementation
is machine independent.

In summary, the interposition agent was more logically structured, was prob-
ably simpler to write and modify, and required no system modifications to im-
plement or run. The kernel-based tracing tools were more efl[icient.

4 Related Work
This section presents a brief survey of past work providing the ability to inter-
pose user code at the system interface or to otherwise extend the functionality
available through the system interface. This topic does not appear to be well
 361

described in the literature; despite intensive research into past systems I have
been unable to find a comprehensive treatment of the subject.
In particular, no general techniques for building or structuring system inter-
face interposition agents appear to have been in use, and so none are described.
Even though a number of systems provided mechanisms by which interposition
agents could be built, the agents that were built appear to have shared little
or no common ground. No widely applicable techniques appear to have been
developed; no literature appears to have been published describing those ad hoc
techniques that were used.
Thus, the following treatment is necessarily somewhat anecdotal in nature,
with some past interposition agents and other system extensions described only
by personal communications. Nonetheless, this section attempts to provide a
representative, if not comprehensive, overview of the related work.

4.1 Overview of Past Agents

A large number of systems have provided low-level facilities sufficient to interpose

user code at the system interface. Both the number and types of interposition
agents that have been built using these facilities have varied widely between the
different systems. Those agents that have been built can be broken down into
five somewhat overlapping categories:

1. Complete operating system emulations such as VM [33] emulating OS/360

or TSO, TENEX [7,48] emulating TOPS-10, and RSEXEC [47] emulating
an Arpanet-wide TENEX.
2. Debuggers and debugging facihties, such as those for (CAL TSS) [43], DTSS
[24], ITS [14], and SunOS [44].
3. System call trace facilities, such as the "SET TRAP JSYS" facility under the
TENEX/Tops-20 [12] EXEC, the t r u s s program under UNIX System V.4
[4], and the t r a c e program under SunOS version 4.
4. Adding new facilities to the operating system interface, as was done in 0S6
[41,42], CAL TSS, and the MS-DOS [29]/Windows [28] environment.
5. Providing enhanced implementations of the existing operating system inter-
face (often enhanced filesystem implementations), as was done in CAL TSS,
TENEX, the Newcastle Connection [8], NFS [50], ITOSS [34], Watchdogs
[6], Taos [26], and particularly in the Macintosh [2] and MS-DOS.

Stackable interfaces are known to be useful in other domains as well. For

instance, communication protocols are composed from stackable layers in the x-
Kernel [20]. Streams are regularly stacked under UNIX System V [4]. Stackable
layers are used for constructing flexible filesystems in Ficus [18,17] and with an
enhanced Vnode interface [11].
Finally, interposition is certainly commonly used on communication channels
in message-based systems. For instance, interposition was regularly used in the
Accent [35] and V [9] systems.
 362

4.2 Analysis of Past Agents

Each of these interposition agents was constructed by hand; almost no code

was reused. In particular, whatever boilerplate code was necessary in order to
intercept, decode and interpret calls made to the raw system interface typically
had to be constructed for each agent. Whatever levels of abstraction that were
necessary in order to build each agent were typically constructed from scratch.
Nonetheless, despite these difficulties, a number of applications of interpo-
sition have been built, taking advantage of the apparent flexibility and config-
urability provided by utilizing a layered approach to system implementation. In
particular, the fact that today people pay real money for interposition agents
that provide enhanced implementations of operating system interfaces (see [31,
37,32,46,28,39,45,49] to name just a few) appears to validate the claim that
interposition can be a useful and effective system building paradigm.

5 Conclusions

5.1 Summary of Results

This research has demonstrated that the system interface can be added to the set
of extensible operating system interfaces that can be extended through interpo-
sition. Just as interposition is successfully used today with such communication-
based facilities as pipes, sockets, and inter-process communication channels, this
work has demonstrated that interposition can be successfully applied to the sys-
tem interface. This work extends the known benefits of interposition to a new
domain.
It achieves this result through the use of an interposition toolkit that substan-
tially increases the ease of interposing user code between clients and instances
of the system interface. It does so by allowing such code to be written in terms
of the high-level objects provided by this interface, rather than in terms of the
intercepted system calls themselves.
The following achievements demonstrate this result:

— an implementation of a system call interposition toolkit for the 4.3BSD in-

terface has been built under Mach,
— the toolkit has been used to construct the agents previously described,
— major portions of the toolkit have been reused in multiple agents,
— the agents have gained leverage by utilizing additional functionality provided
by the toolkit, substantially simplifying their construction, and
— the performance cost of using the toolkit can be small relative to the cost of
the system call interception mechanism and the operations being emulated.

A more detailed presentation of these results can be found in [22].

 363

5.2 Contribution
This research has demonstrated both the feasibihty and the appropriateness of
extending the system interface via interposition. It has shown that while the
4.3BSD system interface is large, it actually contains a small number of abstrac-
tions whose behavior is largely independent. Furthermore, it has demonstrated
that an interposition toolkit can exploit this property of the system interface. In-
terposition agents can both achieve acceptable performance and gain substantial
implementation leverage through use of an interposition toolkit.
These results should be applicable beyond the initial scope of this research.
The interposition toolkit should port to similar systems such as SunOS and
UNIX System V. Agents written for the toolkit should also port. The lessons
learned in building this interposition toolkit should be applicable to building
similar toolkits for dissimilar systems, as explored in [21]. For instance, interpo-
sition toolkits could be constructed for such interfaces as the MS-DOS system
interface, the Macintosh system interface, and the X Window System interface.
Today, agents are regularly written to be interposed on simple communication-
based interfaces such as pipes and sockets. Similarly, the toolkit makes it possible
to easily write agents to be interposed on the system interface. Indeed, it is an-
ticipated that the existence of this toolkit will encourage the writing of such
agents, many of which would not otherwise have been attempted.

5.3 Applicability and Tradeoffs

Interposition is one of many techniques available. As in other domains such as
pipes, filters, IPC intermediaries, and network interposition agents, sometimes
its use will yield a substantial benefit, while sometimes its use would be inap-
propriate. As with other layered techniques, peak achievable performance will
usually not be achieved. Nonetheless, interposition provides a flexibility and ease
of implementation that would not otherwise be available.

5.4 Vision and Potential

The potential opened up by interposition is enormous. Agents can be as easy to
use as filters. They can be as easy to construct as normal application programs.
The system interface can be dynamically customized. Interface changes can be
selectively applied. Indeed, interposition provides a powerful addition to the suite
of application and system building techniques.

Acknowledgments
I'd like to extend special thanks to Rick Rashid, Eric Cooper, M. Satyanarayanan
(Satya), Doug Tygar, Garret Swart, Brian Bershad, and Patricia Jones. Each of
you have have offered helpful suggestions and criticisms which have helped shape
and refine this research.
I'd also like to thank Mark Weiser for his valuable comments on ways to
improve this paper.
 364

References
1. M. Accetta, R. Baxon, D. Golub, R. Rashid, A. Tevanian, and M. Young. Mach:
A new kernel foundation for UNIX development. In Proc. Summer 1986 USENIX
Technical Conference and Exhibition, June 1986.
2. Apple Computer, Inc. Macintosh System Software User's Guide Version 6.0, 1988.
3. AT&T, Customer Information Center, P.O. Box 19901, Indianapolis, IN 46219.
System V Interface Definition, Issue 2, 1986.
4. AT&T. Unix System V Release 4-0 Programmer's Reference Manual, 1989.
5. Robert V. Baron, David Black, William Bolosky, Jonathan Chew, Richard P.
Draves, David B. Golub, Richard F. Rashid, Avadis Tevanian, Jr., and
Michael Wayne Young. Mach Kernel Interface Manual. Carnegie Mellon Uni-
versity School of Computer Science, August 1990.
6. Brian N. Bershad and C. Brian Pinkerton. Watchdogs: Extending the unix iilesys-
tem. In Winter Usenix Conference Proceedings, Dallas, 1988.
7. D. G. Bobrow, J. D. Burchfiel, D. L. Murphy, and R. S. Tomlinson. TENEX,
a paged time sharing system for the PDP-10. Communications of the ACM,
15(3):135-143, March 1972.
8. D. R. Brownbridge, L. F. Marshall, and B. Randell. The Newcastle Connection,
or UNIXes of the world unite! Software - Practice and Experience, 12:1147-1162,
1982.
9. David R. Cheriton, The V distributed system. Communications of the ACM,
31(3):314-333, March 1988.
10. F. W. Clegg, G. S.-F. Ho, S. R. Kusmar, and J. R. Sontag. The HP-UX oper-
ating system on HP Precision Architecture computers. Hewlett-Packard Journal,
37(12):4-22, December 1986.
11. David S. H. Rosenthal. Evolving the vnode interface. In USENIX Conference
Proceedings, pages 107-118. USENIX, June 1990.
12. Digital Equipment Corporation. DECSYSTEM-20 Monitor Calls Reference Man-
ual, January 1978.
13. Digital Equipment Corporation. ULTRIX Reference Pages, Section 2 System Calls,
1989.
14. D. Eastlake, R. Greenblatt, J. HoUoway, T. Knight, and S. Nelson. ITS 1.5 reference
manual. Memorandum no. 161, M.I.T. Artificial IntelHgence Laboratory, July 1969.
Revised form of ITS 1.4 Reference Manual, June 1968.
15. S. I. Feldman. Make - a program for maintaining computer programs. Software -
Practice and Experience, 9(4):255-265, 1979.
16. David Golub, Randall Dean, Alessandro Forin, and Richard Rashid. Unix as an
application program. In Summer Usenix Conference Proceedings, Anaheim, June
1990.
17. Richard G. Guy, John S. Heidemann, Wai Mak, Thomas W. Page, Jr., Gerald J.
Popek, and Dieter Rothmeier. Implementation of the Ficus replicated file system.
In USENIX Conference Proceedings, pages 63-71. USENIX, June 1990.
18. John S. Heidemann. Stackable layers: an architecture for file system development.
Master's thesis, University of California, Los Angeles, July 1991. Available as
UCLA technical report CSD-910056.
19. J. H. Howard, M. L. Kazax, S. G. Menees, D. A. Nichols, M. Satyanarayanan, R. N.
Sidebotham, and M. J. West. Scale and performance in a distributed file system.
ACM TVansactions on Computer Systems, 6(1), February 1988.
 365

20. N. C. Hutchinson and L. L. Peterson. Design of the x-kernel. In Proceedings of the

SIGCOMM '88 Symposium, pages 65-75, Stanford, CA, August 1988.
21. Michael B. Jones. Inheritance in unlikely places: Using objects to build derived
implementations of flat interfaces. In Proceedings of the Third International Work-
shop on Object Orientation in Operating Systems, Paris, September 1992.
22. Michael B. Jones. TVansparently Interposing User Code at the System Interface.
PhD thesis, Carnegie Mellon University, September 1992. Available as Technical
Report CMU-CS-92-170.
23. J.J. Kistler and M. Satyanarayanan. Disconnected operation in the coda file sys-
tem. ACM TYansactions on Computer Systems, 10(1), February 1992.
24. Philip Koch and David Gelhar. DTSS System Programmer's Reference Manual.
Dartmouth College, Hanover, NH, November 1986. Kiewit Computation Center
TM059.
25. Samuel J. Leffler, Marshall Kirk McKusick, Michael J. Karels, and John S. Quar-
terman. The Design and Implementation of the 4.3BSD UNIX Operating System.
Addison-Wesley, October 1990.
26. Paul R. McJones and Garret F. Swart. Evolving the unix system interface to
support multithreaded programs. Research Report 21, Digital Equipment Corpo-
ration, Systems Research Center, September 1987.
27. M. K. McKusick, W. N. Joy, S. J. Leffler, and R. S. Fabry. A fast file system for
unix. ACM Transactions on Computer Systems, 2(3), August 1984.
28. Microsoft Corporation. Microsoft Windows User's Guide, 1987.
29. Microsoft Corporation. Microsoft MS-DOS Operating System version 5.0 User's
Guide and Reference, 1991.
30. Lily B. Mummert and M. Satyanarayanan. Efficient and portable file reference
tracing in a distributed workstation environment. To be published as a Carnegie
Mellon University School of Computer Science technical report, June 1992.
31. Peter Norton Computing, Incorporated. The Norton Utilities for the Macintosh.
1990.
32. Now Software, Inc. Now Utilities: File & Application Management, System Man-
agement, and System Extensions, 1990.
33. R. P. Parmelee, T. I. Peterson, C. C. TiUman, and D. J. Hatfield. Virtual storage
and virtual machine concepts. IBM Systems Journal, 11(2):99-130, 1972.
34. Michael O. Rabin and J. D. Tygar. An integrated toolkit for operating system
security. Technical Report TR-05-87, Harvard University Center for Research in
Computing Technology, Cambridge, MA, May 1987. Revised August 1988.
35. R. F. Rashid and G. Robertson. Accent: A communication oriented network op-
erating system kernel. In Proceedings of the 8th Symposium on Operating Systems
Principles, pages 64-75, December 1981.
36. Brian K. Reid and Janet H. Walker. SCRIBE Introductory User's Manual. UNI-
LOGIC, Ltd., third edition, May 1980.
37. Salient Software, Inc. DiskDoubler User's Manual, 1991.
38. M. Satyanarayanan, J. J. Kistler, P. Kumar, M. E. Okasaki, E. H. Siegel, and
D. C. Steere. Coda: A highly available file system for a distributed workstation
environment. IEEE TYansactions on Computers, 39(4), April 1990.
39. Stac Electronics, Inc. Stacker 2.1 User Guide, 1992.
40. Richard M. Stallman. Using and Porting GNU CC, for version 1.37. Free Software
Foundatation, Inc., 1990.
41. J. E. Stoy and C. Strachey. 0S6 - an experimental operating system for a small
computer. Part 1: General principles and structure. Computer Journal, 15(2):117-
124, May 1972.
 366

42. J. E. Stoy and C. Strachey. 0S6 - an experimental operating system for a small
computer. Part 2: Input/output and filing system. Computer Journal, 15(3):195-
203, August 1972.
43. Howard Ewing Sturgis. A postmortem for a time sharing system. Xerox Research
Report CSL-74-1, Xerox Palo Alto Research Center, January 1974.
44. Sun Microsystems, Inc. SunOS Reference Manual, May 1988. Part No. 800-1751-
10.
45. Symantec Corporation. The Norton Antivirus, 1991.
46. Symantec Corporation. Symantec Antivirus for Macintosh, 1991.
47. Robert H. Thomas. A resource sharing executive for the ARPANET. In Proceedings
of the AFIPS National Computer Conference, volume 42, pages 155-163, June
1973.
48. Robert H. Thomas. JSYS traps - a Tenex mechanism for encapsulation of user
processes. In Proceedings of the AFIPS National Computer Conference, volume 44,
pages 351-360, 1975.
49. Trend Micro Devices, Incorporated. PC-cillin Virus Immune System User's Man-
ual, 1990.
50. D. Walsh, B. Lyon, G. Sager, J. M. Chang, D. Goldberg, S. Kleiman, T. Lyon,
R. Sandberg, and P. Weiss. Overview of the sun network filesystem. In Winter
Usenix Conference Proceedings, Dallas, 1985.
 367

Interposition Agents: An Afterword

Interposition Agents were intended to make it easy to produce custom envi-
ronments in which applications would execute. In 1991, when I first began the
work on Interposition Agents, the idea of virtual machine environments was not
new, although the idea that many of them could be easily constructed to provide
different facilities to or to restrict the behavior of unmodified applications was
new.
Computing has changed a lot in the almost eight years that have passed since
I developed Interposition Agents. Much of this change is due to the Internet. In
this time the Internet has gone from an obscure tool largely only known of and
used by Computer Science researchers and those in the computer industry to a
new mass communication medium.
One example of these changes is that the frequency with which individuals
obtain and use new programs has increased by several orders of magnitude.
Formerly most new programs were distributed as shrink-wrapped applications,
explicitly downloaded shareware, or via copies on floppy disks passed from person
to person (sometimes legally, sometimes not). All required a conscious purchase,
download, or install step to be taken by the user.
Compare that to today's Internet. Web users may download and invoke sev-
eral or even dozens of new programs a day without even being aware that they
are doing so. Programs embedded in web pages may be in numerous languages,
from Java to JavaScript to Visual Basic Script to native code for Intel, PowerPC,
and other machine architectures. Execution environments also vary widely, from
interpreted browser scripts to virtual machine environments with various exe-
cution restrictions, running both interpreted and Just-In-Time compiled code,
to separate processes running with restricted rights, to full native access to the
user's execution state and privileges.
The diversity of execution environments looks likely to continue increasing.
For instance, software fault isolation allows safety properties of native code to
be guaranteed even while executing at near native speeds. Even proof carrying
code is at least a theoretical possibility.
This diversity of execution environments brings us back to Interposition
Agents. The ability to easily construct a multiplicity of execution environments
with different, custom tailored properties, was novel and of some practical use
in 1991. But with the advent of routine transparent downloading and execution
of code from untrusted sources, the need for building customized and secure
environments in which that code executes seems much more compelling today.
Interposition Agents are one technique for doing precisely this.
Providing instances of the system call interface with modified or restricted
properties is one way of executing untrusted code in a safe manner. To be fair,
I will point out that agents built with my particular toolkit, which supported
building agents for Mach 2.5-based UNIX systems, could not enforce additional
security restrictions since agents ran in the same address spaces as the applica-
tions whose execution environments they provided. Nonetheless, as described in
the paper, if other system call interception techniques were used that can send
 368

system calls to distinct processes, such as those provided by Solaris, System V.4,
and some versions of Mach 3.0, agents can make and enforce security guarantees.
I believe the particular interception technique to be largely independent of the
bulk of the techniques used in the Interposition Agents toolkit.
It is my hope that readers reviewing the Interposition Agents work will have
a broadened understanding of the possibilities for easily building a multiplic-
ity of coexisting custom execution environments for application code. With the
rise of applets, controls, scripts and the other forms of executable content, the
poignancy of the need for such environments seems ever more evident.

- Michael B. Jones, Redmond Washington, November 1998

J-Kernel: A Capability-Based Operating System
for Java

Thorsten von Eicken, Chi-Chao Chang, Grzegorz Czajkowski,

Chris Hawblitzel, Deyu Hu, and Dan Spoonhower

Abstract. Safe language technology can be used for protection within

a single address space. This protection is enforced by the language's type
system, which ensures that references to objects cannot be forged. A safe
language alone, however, lacks many features taken for granted in more
traditional operating systems, such as rights revocation, thread protec-
tion, resource management, and support for domain termination. This
paper describes the J-Kernel, a portable Java-based protection system
that addresses these issues. J-Kernel protection domains can commu-
nicate through revocable capabilities, but are prevented from directly
sharing unrevocable object references. A number of micro-benchmaxks
characterize the costs of language-based protection, and an extensible
web and telephony server based on the J-Kernel demonstrates the use of
language-based protection in a large application.

1 Introduction
The notion of moving code across the network to the most appropriate host for
execution has become commonplace. Most often code is moved for efficiency, but
sometimes it is for privacy, for fault-tolerance, or simply for convenience. The
major concern when moving code is security: the integrity of the host to which
it is moved is at risk, as well as the integrity of the computation performed by
the moved code itself.
A number of techniques have been used to place protection boundaries be-
tween so-called "untrusted code" moved to a host and the remainder of the
software running on that host. Traditional operating systems use virtual mem-
ory to enforce protection between processes. A process cannot directly read and
write other processes' memory, and communication between processes requires
traps to the kernel. By limiting the traps an untrusted process can invoke, it
can be isolated to varying degrees from other processes on the host. However,
there's little point in sending a computation to a host if it cannot interact with
other computations there; inter-process communication must be possible.
The major problems encountered when using traditional operating system
facilities to isolate untrusted code are deciding whether a specific kernel trap is
permissible or not and overcoming the cost of inter-process communication. The
semantic level of kernel traps does not generally match the level at which protec-
tion policies are specified when hosting untrusted code. In addition, the objects
on which the traps act are the ones managed by the kernel and not the ones
 370

provided by the hosting environment. Regarding performance, despite a decade

of research leading to a large number of fast inter-process communication mech-
anisms [4, 9, 24], the cost of passing through the kernel and of switching address
spaces remains orders of magnitude larger than that of calling a procedure.
In the context of mobile code, language-based protection is an attractive
alternative to traditional operating system protection mechanisms. Language-
based protection rests on the safety of a language's type system, which ensures
that the abstractions provided by the language's types are enforced. A type
system acts as a simple access control mechanism: it limits the objects that a
computation can access (there is no way to "forge a pointer" to an object), and
it limits the operations that code can perform on accessible objects.
The attraction of language-based protection is twofold: precision of protec-
tion and performance of communication across protection boundaries. Language-
based protection mechanisms allow access rights to be specified with more preci-
sion than traditional virtual-memory based mechanisms: the data items to which
access is permitted as well as the types of accesses permitted can be specified
more finely. For example, in Java, access can be granted precisely to individual
objects and even only to certain object fields using the p u b l i c qualifier. In ad-
dition, with language-based protection, calls across protection boundaries could
potentially be as cheap as simple function calls, enabling as much communication
between components as desired without performance drawbacks.
But language-based protection alone does not make an operating system.
Several projects [1, 2, 3, 8, 11, 14, 35] have recently described how to build
protection domains around components in a safe language environment. The
central idea is to use object references (i.e., pointers to objects) as capabilities for
cross-domain communication. Object references in safe languages are unforgeable
and can thus be used to confer certain rights to the holder(s). In an object-
oriented language, the methods applicable to an object are in essence call gates.
However, this approach, while both fiexible and fast, suffers from two limitations:
there is no way to revoke access to object references, and there is no way to track
which domain owns which objects. This leads to severe problems with domain
termination and resource accounting.
The J-Kernel [16] is an operating system layered on top of a Java Virtual Ma-
chine (JVM) that introduces additional mechanisms borrowed from traditional
operating systems to provide the features missing at the language level. The goal
of the J-Kernel is to define clear boundaries between protection domains, which
are called tasks in the J-Kernel. This makes resource management and task ter-
mination tractable, and simplifies the analysis of inter-task communication. The
boundaries are established by limiting the types of objects that can be shared
between tasks. In the J-Kernel, only special objects called capabilities may be
shared; all other objects are confined to single tasks. This allows the J-Kernel to
equip capabilities with features like revocation, without adding any overheads
to ordinary, non-capability objects.
The main benefits of the J-Kernel are a highly flexible protection model, low
overheads for communication between software components, and operating sys-
 371

tem independence. The current J-Kernel implementation is written entirely in

Java [13] and runs on standard Java Virtual Machines (JVMs). This implementa-
tion was chosen for practical reasons—Java is emerging as the most widely used
general-purpose safe language, and dependable virtual machines are widespread
and easy to work with. While Java in itself does allow for multiple protection
domains within a single JVMs using the sandbox model for applets, that model
is currently very restrictive. It lacks many of the characteristics that are taken
for granted in more traditional systems and, in particular, does not provide a
clear way for different protection domains to communicate with each other.
The J-Kernel also offers a simple form of resource management based on
the JRes interface [7]. The interface allows accounting for resource consumption
on per-thread and per-task basis. The resources accounted for are CPU time,
heap memory, and network traffic. The implementation of JRes contains a small
native component for interfacing to the threads package but is mostly based on
Java bytecode rewriting.
Language-based protection does have drawbacks. First, code written in a
safe language tends to run more slowly than code written in C or assembly
language, and thus the improvement in cross-domain communication may be
offset by an overall slowdown. Much of this slowdown is due to current Java
just-in-time compilers optimized for fast compile times at the expense of run-
time performance. However, even with sophisticated optimization it seems likely
that Java programs will not run as fast as C programs. Second, all current
language-based protection systems are designed around a single language, which
limits developers and doesn't handle legacy code. Software fault isolation [34] and
verification of assembly language [26, 27, 28, 32] may someday offer solutions,
but are still an active area of research.

2 Language-Based P r o t e c t i o n Background

In an unsafe language, any code running in an address space can potentially

modify any memory location in that address space. While, in theory, it is possible
to prove that certain pieces of code only modify a restricted set of memory
locations, in practice this is very difficult for languages like C and arbitrary
assembly language [5, 27], and cannot be fully automated. In contrast, the type
system and the linker in a safe language restrict what operations a particular
piece of code is allowed to perform on which memory locations.
The term namespace can be used to express this restriction: a namespace is
a partial function mapping names of operations to the actions taken when the
operations are executed. For example, the operation "read the field out from the
class System" may perform different actions depending on what class the name
System refers to.
Protection domains around software components can be constructed in a
safe language system by providing a separate namespace for each component.
Communication between components can then be enabled by introducing shar-
ing among namespaces. Java provides three basic mechanisms for controlling
 372

namespaces: selective sharing of object references, static access controls, and

selective class sharing.

Selective Sharing of Object References Two domains can selectively share

references to objects by simply passing each other these references. In the ex-
ample below, raethodl of class A creates two objects of type A, and passes a
reference to the first object to method2 of class B. Since method2 acquires a
reference to a l , it can perform operations on it, such as incrementing the field
j . However, method2 was not given a reference to a2 and thus has no way of
performing any operations on it. Java's safety prevents method2 from forging a
reference to a2, e.g., by casting an integer holding a2's address to a pointer.

class A {
private int i;
public int j;
public static void methodlO {
A al = new A();
A a2 = new A();
B.method2(al);
>

class B {
public static void method2(A arg) {
arg.j++;
}
}

Static Access Control The preceding example demonstrated a very dynamic

form of protection—methods can only perform operations on objects to which
they have been given a reference. Java also provides static protection mechanisms
that limit what operations a method can perform on an object once the method
has acquired a reference to that object. A small set of modifiers can change
the scope of fields and methods of an object. The two most common modifiers,
p r i v a t e and p u b l i c , respectively limit access to methods in the same class or
allow access to methods in any class. In the classes shown above, method2 can
access the public field j of the object a l , but not the private field i.

Selective Class Sharing Domains can also protect themselves through control
of their class namespace. To understand this, we need to look at Java's class
loading mechanisms. To allow dynamic code loading, Java supports user-defined
class loaders which load new classes into the virtual machine at run-time. A class
loader fetches Java bytecode from some location, such as a file system or a URL,
and submits the bytecode to the virtual machine. The virtual machine performs
a verification check to make sure that the bytecode is legal, and then integrates
 373

the new class into the machine execution. If the bytecode contains references to
other classes, the class loader is invoked recursively in order to load those classes
as well.
Class loaders can enforce protection by making some classes visible to a
domain, while hiding others. For instance, the example above assumed that
classes A and B were visible to each other. However, if class A were hidden from
class B (i.e. it did not appear in B's class namespace), then even if B obtains a
reference to an object of type A, it will not be able to access the fields i and j ,
despite the fact that j is pubhc.

2.1 Straight-Forward Protection Domains: The Share Anything

Approach

The simple controls over the namespace provided in Java can be used to construct
software components that communicate with each other but are still protected
from one another. In essence, each component is launched in its own namespace,
and can then share any class and any object with other components using the
mechanisms described above. While we will continue to use the term protection
domain informally to refer to these protected components, we will axgue that it
is impossible to precisely define protection domains when using this approach.
The example below shows a hypothetical file system component that gives
objects of type FileSystemlnterf ace to its clients to give them access to files.
Client domains make cross-domain invocations on the file system by invoking the
open method of a FileSystemlnterf ace object. By specifying different values
for accessRights and r o o t D i r e c t o r y in different objects, the file system can
enforce different protection policies for different clients. Static access control en-
sures that clients cannot modify the accessRights and r o o t D i r e c t o r y fields di-
rectly, and one client cannot forge a reference to another client's FileSystemlnterf ace
object.

class FileSystemlnterface {
p r i v a t e i n t accessRights;
p r i v a t e Directory r o o t D i r e c t o r y ;
p u b l i c F i l e open(String fileName) {}
}

The filesystem example illustrates an approach to protection in Java that re-

sembles a capability system. Several things should be noted about this approach.
First, this approach does not require any extensions to the Java language—all
the necessary mechanisms already exist. Second, there is very little overhead
involved in making a call from one protection domain to another, since a cross-
domain call is simply a method invocation, and large arguments can be passed
by reference, rather than by copy. Third, references to any object may be shared
between domains since the Java language has no way of restricting which ref-
erences can be passed through a cross-domain method invocation and which
cannot.
 374

When we first began to explore protection in Java, this share anything ap-
proach seemed the natural basis for a protection system, and we began develop-
ing on this foundation. However, as we worked with this approach a number of
problems became apparent.

Revocation The first problem is that access to an object reference cannot be

revoked. Once a domain has a reference to an object, it can hold on to it forever.
Revocation is important in enforcing the principle of least privilege: without
revocation, a domain can hold onto a resource for much longer than it actually
needs it.
The most straightforward implementation of revocation uses extra indirec-
tion. The example below shows how a revocable version of the earlier class A
can be created. Each object of A is wrapped with an object of AWrapper, which
permits access to the wrapped object only until the revoked flag is set.

class A {
p u b l i c i n t m e t h K i n t a l , i n t a2) {}
}

c l a s s AWrapper {
p r i v a t e A a;
p r i v a t e boolean revoked;
p u b l i c i n t m e t h l ( i n t a l , i n t a2) {
i f ( ! r e v o k e d ) r e t u r n a . m e t h l ( a l , a2);
e l s e throw new RevokedExceptionO;
}
p u b l i c void revoke() { revoked=true; }
p u b l i c AWrapper(A realA) {
a = realA; revoked = f a l s e ; }
}
In principle, this solves the revocation problem and is efficient enough for
most purposes. However, our experience shows that programmers often forget to
wrap an object when passing it to another domain. In particular, while it is easy
to remember to wrap objects passed as arguments, it is common to forget to wrap
other objects to which the first one points. In effect, the default programming
model ends up being an unsafe model where objects cannot be revoked. This
is the opposite of the desired model: safe by default and unsafe only in special
cases.

Inter-domain Dependencies a n d Side Effects As more and more object

references are shared between domains, the structure of the protection domains
is blurred, because it is unclear from which domains a shared object can be
accessed. For the programmer, it becomes difficult to track which objects are
shared between protection domains and which are not, and the Java language
 375

provides no help as it makes no distinction between the two. Yet, the distinction
is critical for reasoning about the behavior of a program running in a domain.
Mutable shared objects can be modified at any time in by other domains that
have access to the object, and a programmer needs to be aware of this possible
activity. For example, a malicious user might try to pass a byte array holding
legal bytecode to a class loader (byte arrays, like other objects, are passed by
reference to method invocations), wait for the class loader to verify that the
bytecode is legal, and then overwrite the legal bytecode with illegal bytecode
which would subsequently be executed. The only way the class loader can protect
itself from such an attack is to make its own private copy of the bytecode, which
is not shared with the user and is therefore safe from malicious modification.

Domain Termination The problems associated with shared object references

come to a head when we consider what happens when a domain must be ter-
minated. Should all the objects that the domain allocated be released, so that
the domain's memory is freed up? Or should objects allocated by the domain
be kept alive as long as other domains still hold references to them? ^Prom a
traditional operating systems perspective, it seems natural that when a process
terminates all of its objects disappear, because the address space holding those
objects ceases to exist. On the other hand, from a Java perspective, objects can
only be deallocated when there are no more reachable references to them.
Either solution to domain termination leads to problems. Deallocating ob-
jects when the domain terminates can be extremely disruptive if objects are
shared at a fine-grained level and there is no explicit distinction between shared
and non-shared objects. For example, consider a Java S t r i n g object, which
holds an internal reference to a character array object. Suppose domain 2 holds
a S t r i n g object whose internal character array belongs to domain 1. If domain
1 dies, then the S t r i n g will suddenly stop working, and it may be beyond the
programmer's ability to deal with disruptions at this level.
On the other hand, if a domain's objects do not disappear when the domain
terminates, other problems can arise. First, if a server domain fails, its clients
may continue to hold on to the server's objects and attempt to continue using
them. In effect, the server's failure is not propagated correctly to the clients.
Second, if a client domain holds on to a server's objects, it may indirectly also
hold on to other resources, such as open network connections and files. A careful
server implementation could explicitly relinquish important resources before ex-
iting, but in the case of unexpected termination this may be impossible. Third,
if one domain holds on to another domain's objects after the latter exits, then
any memory leaks in the terminated domain may be unintentionally transferred
to the remaining one. It is easy to imagine scenarios where recovery from this
sort of shared memory leak requires a shutdown of the entire VM.

Threads By simply using method invocation for cross-domain calls, the caller
and callee both execute in the same thread, which creates several potential haz-
ards. First, the caller must block until the callee returns—there is no way for the
 376

caller to gracefully back out of the call without disrupting the callee's execution.
Second, Java threads support methods such as stop, suspend, and s e t P r i o r i t y
that modify the state of a thread. A malicious domain could call another do-
main and then suspend the thread so that the callee's execution gets blocked,
perhaps while holding a critical lock or other resource. Conversely, a malicious
callee could hold on to a Thread object and modify the state of the thread after
execution returns to the caller.

Resource Accounting A final problem with the simple protection domains

is that object sharing makes it difficult to hold domains accountable for the
resources that they use, such as processor time and memory. In particular, it is
not clear how to define a domain's memory usage when domains share objects.
One definition is that a domain is held accountable for all of the objects that it
allocates, for as long as those objects remain alive. However, if shared objects
aren't deallocated when the domain exits, a domain might continue to be charged
for shared objects that it allocated, long after it has exited. Perhaps the cost of
shared objects should be split between all the domains that have references to
the object. However, because objects can contain references to other objects, a
malicious domain could share an object that looks small, but actually contains
pointers to other large objects, so that other domains end up being charged for
most of the resources consumed by the malicious domain.

Summary The simple approach to protection in Java outlined in this section is

both fast and flexible, but it runs into trouble because of its lack of structure. In
particular, it fails to clearly distinguish between the ordinary, non-shared object
references that constitute a domain's internal state, and the shared object refer-
ences that are used for cross-domain communication. Nevertheless, this approach
is useful to examine, because it illustrates how much protection is possible with
the mechanisms provided by the Java language itself. It suggests that the most
natural approach to building a protection system in Java is to make good use
of the language's inherent protection mechanisms, but to introduce additional
structure to fix the problems. The next section presents a system that retains
the flavor of the simple approach, but makes a stronger distinction between
non-shared and shared objects.

3 The J-Kernel
The J-Kernel is a capability-based system that supports multiple, cooperating
protection domains, called tasks, which run inside a single Java virtual machine.
Capabilities were chosen because they have several advantages over access control
lists: (i) they can be implemented naturally in a safe language, (ii) they can
enforce the principle of least privilege more easily, and (iii) by avoiding access
list lookups, operations on capabilities can execute quickly.
The primary goals of the J-Kernel are:
 377

- a precise definition of tasks, with a clear distinction between objects local to

a task and capability objects that can be shared between tasks,
- well defined, flexible communication channels between tasks based on capa-
bilities,
- support for revocation for all capabilities, and
- clean semantics of task termination.

To achieve these goals, we were willing to accept higher cross-task commu-

nication overheads when compared to the share anything approach. In order to
ensure portability, the J-Kernel is implemented entirely as a Java library and
requires no native code or modifications to the virtual machine. To accomplish
this, the J-Kernel defines a class loader that examines and in some cases modifies
user-submitted bytecode before passing it on to the virtual machine. This class
loader also generates bytecode at run-time for stub classes used for cross-task
communication. Finally, the J-Kernel's class loader substitutes safe versions for
some problematic standard classes. With these implementation techniques, the
J-Kernel builds a protection architecture that is radically different from the se-
curity manager based protection architecture that is the default model on most
Java virtual machines.
Protection in the J-Kernel is based on three core concepts—capabilities,
tasks, and cross-task calls:

- Capabilities are implemented as objects of the class C a p a b i l i t y and repre-

sent handles onto resources in other tasks. A capability can be revoked at
any time by the task that created it. All uses of a revoked capability throw
an exception, ensuring the correct propagation of failure.
- Tasks are represented by the Java class Task. Each task has a namespace
that it controls as well as a set of threads. When a task terminates, all of
the capabilities that it created are revoked, so that all of its memory may be
freed, thus avoiding the task termination problems that plagued the share
anything approach.
- Cross-task calls are performed by invoking methods of capabilities obtained
from other tasks. The J-Kernel's class loader interposes a special calling con-
vention ^ for these calls: arguments and return values are passed by reference
if they are also capabilities, but they are passed by copy if they are primitive
types or non-capability objects. When an object is copied, these rules are
applied recursively to the data in the object's fields, so that a deep copy of
the object is made. The effect is that only capabilities can be shared between
tasks and references to regular objects are confined to single tasks.

3.1 J-Kernel Implementation

The J-Kernel's implementation of capabilities and cross-task calls relies heavily
on Java's interface classes. An interface class defines a set of method signa-
' The standard Java calling convention passes primitive data types (int, float, etc.) by
copy and object data types by reference.
 378

tures without providing their implementation. Other classes that provide cor-
responding implementations can then be declared to implement the interface.
Normally interface classes are used to provide a limited form of multiple in-
heritance (properly called interface inheritance) in that a class can implement
multiple interfaces. In addition, Sun's remote method invocation (RMI) speci-
fication [19] "pioneered" the use of interfaces as compiler annotations. Instead
of using a separate interface definition language (IDL), the RMI specification
simply uses interface classes that are flagged to the RMI system in that they
extend the class Remote. Extending Remote has no effect other than directing
the RMI system to generate appropriate stubs and marshalling code.

Because of the similarity of the J-Kernel's cross-task calls to remote method

invocations, we have integrated much of Sun's RMI specification into the capa-
bility interface. The example below shows a simple remote interface and a class
that implements this remote interface, both written in accordance with Sun's
RMI specification.

/ / i n t e r f a c e c l a s s shared with other t a s k s

i n t e r f a c e ReadFile extends Remote {
byte readByteO throws RemoteException;
b y t e [ ] r e a d B y t e s ( i n t nBytes)
throws RemoteException;
>

/ / implementation hidden from other t a s k s

c l a s s ReadFilelmpl implements ReadFile {
p u b l i c byte readByteO {••.}•
p u b l i c b y t e [ ] r e a d B y t e s ( i n t nBytes) •[. . . }

To create a capability in the J-Kernel, a task calls the c r e a t e method of the

class C a p a b i l i t y , passing as an argument a target object that implements one
or more remote interfaces. The c r e a t e method returns a new capability, which
extends the class Capability and implements all of the remote interfaces that
the target object implements. The capability can then be passed to other tasks,
which can cast it to one of its remote interfaces, and invoke the methods this
interface declares. In the example below task 1 creates a capability and adds it
to the system-wide repository (the repository is a name service allowing tasks
to publish capabilities). Task 2 retrieves the capability from the repository, and
makes a cross-task invocation on it.
 379

// Task 1:
// instantiate new ReadFilelmpl object
ReadFilelmpl target = new ReadFilelmpl();
// create a capability for the new object
Capability c = Capability.create(target);
// add it to repository under some naime
Task.getRepositoryO.bind(
"TasklReadFile", c ) ;

// Task 2:
/ / extract capability
C a p a b i l i t y c = Task.getRepositoryO
.lookup("TasklReadFile");
/ / c a s t i t t o ReadFile, and invoke remote method
byte b = ((ReadFile) c ) . r e a d B y t e O ;

Essentially, a capability object is a wrapper object around the original target

object. The code for each method in the wrapper switches to the task that
created the capability, makes copies of all non-capability arguments according
to the special calling convention, and then invokes the corresponding method
in the target object. When the target object's method returns, the wrapper
switches back to the caller task, makes a copy of the return value if it is not a
capability, and returns.

L o c a l - R M I S t u b s The simple looking call to C a p a b i l i t y . c r e a t e in fact hides

most of the complexity of traditional RPC systems. Internally, c r e a t e automat-
ically generates a stub class at run-time for each target class. This avoids off-line
stub generators and IDL files, and it allows the J-Kernel to speciahze the stubs
to invoke the target methods with minimal overhead. Besides switching tasks,
stubs have three roles: copying arguments, supporting revocation, and protecting
threads.
By default, the J-Kernel uses Java's built-in serialization features [19] to
copy an argument: the J-Kernel serializes an argument into an array of bytes,
and then deserializes the byte array to produce a fresh copy of the argument.
While this is convenient because many built-in Java classes are serializable, it
involves a substantial overhead. Therefore, the J-Kernel also provides a fa.st copy
mechanism, which makes direct copies of objects and their fields without using an
intermediate byte array. The fast copy implementation automatically generates
specialized copy code for each class that the user declares to be a fast copy class.
For cyclic or directed graph data structures, a user can request that the fast
copy code use a hash table to track object copying, so that objects in the data
structure are not copied more than once (this slows down copying, though, so
by default the copy code does not use a hash table).
Each generated stub contains a revoke method that sets the internal pointer
to the target object to n u l l . Thus all capabilities can be revoked and doing so
makes the target object eligible for garbage collection, regardless of how many
 380

other tasks hold a reference to the capabihty. This prevents tasks from holding
on to garbage in other tasks.
In order to protect the caller's and callee's threads from each other, the
generated stubs provide the illusion of switching threads. Because most virtual
machines map Java threads directly onto kernel threads it is not practical to
actually switch threads: as shown in the next subsection this would slow down
cross-task calls substantially. A fast user-level threads package might solve this
problem, but would require modifications to the virtual machine, and would thus
limit the J-Kernel's portability. The compromise struck in the current implemen-
tation uses a single Java thread for both the caller and callee but prevents direct
access to that thread to avoid security problems.
Conceptually, the J-Kernel divides each Java thread into multiple segments,
one for each side of a cross-task call. The J-Kernel class loader then hides the
system Thread class that manipulates Java threads, and interposes its own with
an identical interface but an implementation that only acts on the local thread
segment. Thread modification methods such as stop and suspend act on thread
segments rather than Java threads, which prevents the caller from modifying
the callee's thread segment and vice-versa. This provides the illusion of thread-
switching cross-task calls, without the overhead for actually switching threads.
The illusion is not totally convincing, however—cross-task calls really do block,
so there is no way for the caller to gracefully back out of one if the callee doesn't
return.

Class N a m e Resolvers In the standard Java applet architecture, applets have

very little access to Java's class loading facilities. In contrast, J-Kernel tasks are
given considerable control over their own class loading. Each task has its own
class namespace that maps names to classes. Classes may be local to a task, in
which case they are only visible in that task's namespace or they may be shared
between multiple tasks, in which case they are visible in many namespaces. A
task's namespace is controlled by a user-defined resolver, which is queried by the
J-Kernel whenever a new class name is encountered. A task can use a resolver to
load new bytecode into the system, or it can make use of existing shared classes.
After a task has loaded new classes into the system, it can share these classes
with other tasks if it wants, by making a SharedClass capability available to
other tasks ^.
Shared classes are the basis for cross-task communication: tasks must share
remote interfaces and fast copy classes to establish common methods and argu-
ment types for cross-task calls. Allowing user-defined shared classes makes the
cross-task communication architecture extensible; standard Java security archi-
tectures only allow pre-defined "system classes" to be shared between tasks, and
thus limit the expressiveness of cross-task communication.

^ Shared classes (and, transitively, the classes that shared classes refer to) are not
allowed to have static fields, to prevent sharing of non-capability objects through
static fields. In addition, to ensure consistency between domains, two domains that
share a class must also share other classes referenced by that class.
 381

Ironically, the J-Kernel needs to prevent the sharing of some system classes.
For example, the file system and thread classes present security problems. Others
contain resources that need to be defined on a per-task basis: the class System,
for example, contains static fields holding the standard input/output streams.
In other words, the "one size fits all" approach to class sharing in most Java
security models is simply not adequate, and a more flexible model is essential to
make the J-Kernel safe, extensible, and fast.
In general, the J-Kernel tries to minimize the number of system classes visible
to tasks. Classes that would normally be loaded as system classes (such as classes
containing native code) are usually loaded into a privileged task in the J-Kernel,
and are accessed through cross-task communication, rather than through direct
calls to system classes. For instance, we have developed a task for file system
access that is called using cross-task communication. To keep compatibility with
the standard Java file API, we have also written alternate versions of Java's
standard file classes, which are just stubs that make the necessary cross-task
calls. (This is similar to the interposition proposed by [35]).
The J-Kernel moves functionality out of the system classes and into tasks
for the same reasons that micro-kernels move functionality out of the operating
system kernel. It makes the system as a whole extensible, i.e., it is easy for any
task to provide alternate implementations of most classes that would normally
be system classes (such as file, network, and thread classes). It also means that
each such service can implement its own security policy. In general, it leads to a
cleaner overall system structure, by enforcing a clear separation between differ-
ent modules. Java libraries installed as system classes often have undocumented
and unpredictable dependencies on one another ^. Richard Rashid warned that
the UNIX kernel had "become a 'dumping ground' for every new feature or fa-
cility" [30]; it seems that the Java system classes are becoming a similar dumping
ground.

3.2 J-Kernel Micro-Benchmarks

To evaluate the performance of the J-Kernel mechanisms we measured a number
of micro-benchmarks on the J-Kernel as well as on a number of reference systems.
Unless otherwise indicated, all micro-benchmarks were run on 200Mhz Pentium-
Pro systems running Windows NT 4.0 and the Java virtual machines used were
Microsoft's VM (MS-VM) and Sun's VM with Symantec's JIT compiler (Sun-
VM). All numbers are averaged over a large number of iterations.

Null LRMI Table 1 dissects the cost of a null cross-task call (null LRMI)
and compares it to the cost of a regular method invocation, which takes a few
tens of nanoseconds. The J-Kernel null LRMI takes 60x to 180x longer than
^ For instance, Microsoft's implementation of java.io.File depends on
java.io.DatalnputStream, which depends on com.ms.lang.SystemX, which de-
pends on classes in the abstract windowing toolkit. Similarly, java.lang.Object
depends transitively on almost every standard library class in the system.
 382

a regular method invocation. With MS-VM, a significant fraction of the cost

Ues in the interface method invocation necessary to enter the stub. Additional
overheads include the synchronization cost when changing thread segments (two
lock acquire/release pairs per call) and the overhead of looking up the current
thread. Overall, these three operations account for about 70% of the cross-task
call on MS-VM and about 80% on Sun-VM. Given that the implementations of
the three operations are independent, we expect better performance in a system
that includes the best of both VMs.

Operation MS-VM Sun-VM

Regular Method invocation 0.04^s 0.03^s
Interface method invocation 0.54/is 0.05/xs
Thread info lookup 0.55/is 0.29/^s
Acquire/release lock 0.20/iS 1.91/is
J-Kernel LRMI 2.22//S 5.41/xs
Table 1. Cost of null method invocations

To provide a comparison of the J-Kernel LRMI to traditional OS cross-task

calls. Table 2 shows the cost of several forms of local RPC available on NT.
NT-RPC'is the standard, user-level RPC facility. COM out-of-proc is the cost of
a null interface invocation to a COM component located in a separate process on
the same machine. The communication between two fully protected components
is at least a factor of 3000 from a regular C-l-4- invocation (shown as COM
in-proc). For a comparison to the local RPC performance of research operating
systems see the Related Work section.

Form of RPC Time

NT-RFC 109/is
COM out-of-proc 99/is
COM in-proc O.OS/is
Table 2. Local RPC costs using NT mechanisms

Threads Table 3 shows the cost of switching back and forth between two Java
threads in MS-VM and Sun-VM. The base cost of two context switches between
NT kernel threads {NT-base) is S.Gfis, and Java introduces an additional 1-2/is
of overhead. This confirms that switching Java threads during cross-task calls
would add a significant cost to J-Kernel LRML

Argument Copying Table 4 compares the cost of copying arguments during

a J-Kernel LRMI using Java serialization and using the J-Kernel's fast-copy
 383

NT-base MS-VM Sun-VM

8.6^s 9.8/is 10.2;is
Table 3. Cost of a double thread switch using regular Java threads

mechanism. By making direct copies of the objects and their fields without
using an intermediate Java byte-array, the fast-copy mechanism improves the
performance of LRMI substantially—more than an order of magnitude for large
arguments. The performance difference between the second and third rows (both
copy the same number of bytes) is due to the cost of object allocation and
invocations of the copying routine for every object.

N u m b e r x size MS -VM Su]i - V M

of objects Serializ. Fast-copy Serializ. Fast-Copy
1 X 10 bytes 104/iS 4.8/Lts 331^8 13.7ps
1 X 100 bytes 158;is 7.7/iS 509/is 18.5/is
10 X 10 bytes 193/is 23.3/is 521/xs 79.3/iS
1 X 1000 bytes 633/is 19.2/is 2105//S 66.7//S
Table 4. Cost of Argument Copying during LRMI

In summary, the micro-benchmark results are encouraging in that the cost of

a cross-task call is 50x lower in the J-Kernel than in NT. However, the J-Kernel
cross-task call still incurs a stiff penalty over a plain method invocation.
Inspection of the critical code paths shows that in the call path itself, the
acquisition of a lock and the management of the thread state contribute most to
the cost. While aggressive inlining could reduce that by perhaps a factor of 2x,
the high cost of lock operations in modern processors will keep the cost signif-
icant. The largest cost of most cross-task calls is likely to be in the copying of
the arguments. For small objects the allocation (and eventual garbage collection)
dominates the cost and appears difficult to optimize by more than a small factor.
A more promising avenue would be to optimize the passing of arrays of primitive
(i.e., non-pointer) types by simply sharing the array. This clearly changes the
semantics, but does not compromise the integrity of the J-Kernel.

4 A n Extensible Web and Telephony Server

One of the driving applications for the J-Kernel is an extensible web and tele-
phony server. The goal is to allow users to dynamically extend the functionality
of the server by uploading Java programs, called servlets [21], that customize the
processing of HTTP requests and telephone calls. Applications that integrate the
two forms of communication are of particular interest.
The extensible web and telephony server is made possible by the ongoing
merger of networking and telephony. The resulting new wave of commodity tele-
 384

phony equipment is designed to be controlled by a standard workstation. Exam-

ples are telephone interface cards with 2 to 24 analog telephone lines and the
DSPs to handle real-time tasks, and full telephone switches (PBXs) that can
be controlled over a LAN. On the software side, standard telephony APIs (e.g.,
Windows' TAPI or Java's JTAPI) allow programs to control the equipment in a
relatively portable way.
The upshot of these developments is that today a single commodity PC
can provide integrated communication services that use both the Internet and
the telephone system. This section describes a prototype integrated web and
telephony server called the J-Server, which uses the J-Kernel to support the
extensibility necessary in this setting. The J-Server core manages a hierarchical
namespace of resources where each resource can correspond either to a URL or
to a telephone number. Users of the system can upload programs (called servlets)
to the J-Server and attach each servlet to one or more resources. The servlets can
then handle all HTTP requests and telephone calls that are sent to the resources.
Furthermore, servlets may communicate with one another, so that one servlet
may provide a service to another servlet. For example, a servlet that uses speech
recognition to determine the nature of a call might need to dispatch the call to
another servlet once it is clear whom the call is for.
In general, our experience with using these servers has shown that they place
high demands on the extension mechanism:

— Protection is important because multiple users create their own independent

webs.
— Failure isolation is important because webs are continuously expanded and
new features must not disrupt old ones.
— Clean servlet termination is essential,
— Extensions must be able to call other extensions, and these cross-task calls
should be cheap so that they can occur frequently.
— Resource management and accounting are necessary.

Our J-Server setup uses two off-the-shelf products: Lucent Technologies DEFIN-
ITY Enterprise Communications Server and the Dialogic Dialog/4 voice inter-
face. At the core, the DEFINITY server is a Private Branch Exchange (PBX),
i.e. a telephone switch, augmented by an Ethernet-based network controller. The
network controller in this PBX allows the routing of telephone calls to be moni-
tored and controlled from a PC over the LAN. Lucent provides a Java interface
allowing the J-Server to communicate with the PBX using the Java Telephony
API (JTAPI) [18].
The Dialog/4 Voice Processing Board, from Dialogic Corporation, provides
four half-duplex, analog telephone interfaces and includes a digital signal proces-
sor (DSP) to play and record audio, and to detect and transmit DTMF (dual-
tone multi-frequency) signals. Using two Dialog/4 boards, the J-Server is able to
make eight simultaneous half-duplex voice connections to the telephone network.
 385

4.1 Extensible Server Architecture

Each servlet in the J-Server runs as a separate task to isolate it from the J-
Server core as well as from other servlets. This setup allows new servlets to be
introduced into the system or crashed ones to be reloaded without disturbing the
server operation. The J-Server core is also divided into several tasks. One task
dispatches incoming HTTP requests to servlets. Three tasks handle telephony:
the first is responsible for communicating with the PBX via JTAPI, the second
deals with the Dialogic voice processing boards, and the third is in charge of
dispatching telephony events to the appropriate servlets.
During operation, the J-Server performs a large number of cross-task calls to
pass events and data around. This is best illustrated by looking at the handling
of a telephone call by a simple voice-mail servlet. When a telephone call reaches
the PBX, an event is sent to the J-Server via JTAPI and is handled by the
task in charge of PBX events. This task then sends the event information to the
telephony dispatch task, which in turn passes the information to the appropriate
servlet. The servlet instructs the J-Server to route the call to one of the telephone
lines connected to the Dialogic card. At this point, the voice processing task
begins generating events, starting with a "ringing" event. These events are passed
to the dispatching task, and then on to the appropriate servlet. The servlet
proceeds to take the telephone off hook, and to start playing audio data. In
this application, the data is passed from the servlet task to the dispatch task
and then on to the voice task in one large chunk. When the voice channel has
finished playing the audio sample, it alerts the servlet, which, in turn, responds
by creating an appropriate response object. When the calling party hangs up,
the servlet instructs the voice channel to stop recording. The servlet waits for
all recorded data and then saves the sample as a Microsoft WAVE file for later
playback.

4.2 Performance

We conducted a set of experiments on the J-Server to evaluate the performance

of the J-Kernel in an application setting. The experiments were carried out on a
300MHz Pentium II, running Windows NT 4.0. The Java virtual machine used
was Microsoft's VM version 1.1, with SDK 2.01.
Two tests measure the costs of cross-task calls and the frequency of such
calls in the J-Server. The first test, Half-Duplex Conversation, mimics a 40-
second telephone conversation between two people, each of whom is simulated
by a servlet. The call initiator "speaks" for two seconds, then "listens" for two
seconds, then "speaks" again, while the receiver is doing the opposite. "Speak-
ing" involves sending uncompressed audio data in one kilobyte segments (which
is equivalent to l/8sec of conversation) to the voice processing task, where they
are presented to the Dialogic driver, which plays the data over the phone line.
"Listening" consists of recording data (in 512byte segments), collecting the data
to accumulate two seconds worth of audio (16Kbytes) and writing it out to the
disk in a standard, uncompressed WAVE file format.
 386

The second test, VoiceBox Listing, measures the performance of the HTTP
component of J-Server. Using a web browser, a user makes a request to the
servlet, which displays the contents of a voice box directory, formatted as an
HTML page. Cross-task calls are made from the HTTP dispatching task to the
servlet task. In addition, a simple authentication servlet is used to verify the
user's password, resulting in another cross-task call.
Table 5 summarizes the results. The Initiator and Responder columns cor-
respond to two parties involved in HalfDuplex Conversation, while the Listing
column corresponds to the servlet fetching and formatting information about
voice messages. The first five rows present the actual measurement values; the
remaining rows contain derived information.

Half-duplex conversation HTTP

Initiator Responder Listing
Total CPU time 1503 ms 1062 ms 18.5 ms
Total number of cross-task calls 3085 2671 8
Total cross-task data transfer 962304 bytes 984068 bytes 3596 bytes
Total cross-task call overhead 9.37 ms 8.71 ms 0.037 ms
Total cross-task data copy time 2.91 ms 2.58 ms 0.01 ms
Average CPU time/cross-task call 3.0/iS 3.3 /iS 4.6 /JLS
Average bytes/call 312 bytes 368 bytes 450 bytes
Average copy overhead/call 31% 30% 27%
Average copy bandwidth 331 MB/s 381 MB/s 360 MB/s
Relative cross-task call overhead 0.62 % 0.82 % 0.20 %
Table 5. Selected performance measurements of HalfDuplex Conversation and Voice-
Box Listing.

In all cases the overheads of crossing tasks (which include the cost of copying
arguments and then return values) are below 1% of the total consumed CPU
time. On the average, crossing tasks costs between 3//s-4.6/is and an average
call transfers between 312 (Initiator) to 450 (Listing) bytes. The cost of copying
data accounts for 27%-31% of an average cross-task call. This suggests that
overheads of crossing tasks in real applications when no data transfer is involved
are between 2.1/is-3.3^s. This is in contrast to the cost of a null cross-task call,
measured in an isolated tight loop: 1.35/is.
The achieved bandwidth of copying data across tasks is sufficient for this ap-
plication. About 85% of the data transferred across tasks in the two experiments
is stored in byte arrays of 0.5Kbytes and IKbytes; the remaining data are either
references to capabilities, primitive data types or data inside objects of the type
J a v a . l a n g . S t r i n g . The average copying bandwidth is 330-380Mbytes/s, which
is over 50% of the peak data copying bandwidth of 630Mbytes/s achieved in
Java with a tight loop of calls to System.arraycopyO.
With respect to scalability, our current hardware configuration supports at
most eight simultaneous half-duplex connections. Every connection results in an
 387

additional 1.7%-2.3% increase in the CPU load. From this we estimate that,
given enough physical lines, our system could support about 50 simultaneous
half-duplex connections before reaching full CPU utilization. Since the J-Server
demands modest amounts of physical memory, permanent storage and network
resources, CPU is the bottleneck from the perspective of scaling the system,
so the range of 50 simultaneous half-duplex connections is very likely to be
achievable in practice.
While analyzing the performance, it is important to note that Java introduces
non-obvious overheads. For instance, in the Listing experiment, a large part of
the 18.5ms of CPU time can be accounted for as follows. About seven millisec-
onds are spent in the network and file system Java classes and native protocol
stacks. Roughly eight milliseconds are spent performing string concatenations
in one particular part of the VoiceBox Listing servlet (Java string processing is
very slow under MS JVM). These overheads, which are introduced by Java and
not easy to avoid, dwarf the relative costs of crossing task boundaries and lead to
conclusions that may be not true if the JVM performance improves dramatically.
One conclusion, based on the current performance numbers, is that in large
applications similar to J-Server, the task boundaries crossing overheads are
very small relative to the total execution time and further optimizing them
will bring barely noticeable performance improvements. Second, in applications
where crossing tasks is frequent and optimizing cross-task calls actually matters,
on the average 30% of task crossing overheads can be removed by wrapping data
structures and arrays in capabilities. This avoids copying data between tasks but
comes at the expense of increased programming effort. It is important to stress
that the conclusions are drawn from observing several experiments based on a
single, although complex and realistic, application. More study is needed to fully
understand application behavior on top of Java in general and the J-Kernel in
particular.

5 Related Work
The Alta and GVM systems [1] developed at the University of Utah are closely
related to the J-Kernel. The motivation is to implement a process model in
Java, which provides memory protection, control over scheduling, and resource
management. Both systems are implemented by modifying a JVM (the free
Kaffe JVM is used). In GVM each Java process receives separate threads, heap,
and classes. Sharing is only allowed between a process and a special system
heap. This invariant is enforced using a write barrier. Each heap is garbage
collected separately and a simple reference counting mechanism is used for cross-
heap references. Alta implements the nested process model used in the Fluke
microkernel [10]. Child processes of the same parent process can share data
structures directly: the resources used are attributed to the parent. While the
inter-process communication costs of Alta and GVM are higher than in the J-
Kernel (see [1]), the custom JVM modifications allow both GVM and Alta to
take control of thread scheduling and to be more precise in resource management.
 388

Several major vendors have proposed extensions to the basic Java sandbox se-
curity model for applets [20, 25, 29]. For instance, Sun's JDK 1.1 added a notion
of authentication, based on code signing, while the JDK 1.2 adds a richer struc-
ture for authorization, including classes that represent permissions and methods
that perform access control checks based on stack introspection [12]. JDK 1.2
"protection domains" are implicitly created based on the origin of the code, and
on its signature. This definition of a protection domain is closer to a user in
Unix, while the J-Kernel's task is more like a process in Unix. Balfanz et al. [2]
define an extension to the JDK which associates domains with users running
particular code, so that a domain becomes more like a process. However, if do-
mains are able to share objects directly, revocation, resource management, and
domain termination still need to be addressed in the JDK.
JDK 1.2 system classes are still lumped into a monolithic "system domain",
but a new classpath facilitates loading local applications with class loaders rather
than as system classes. However, only system classes may be shared between
domains that have different class loaders, which limits the expressiveness of
communication between domains. In contrast, the J-Kernel allows tasks to share
classes without requiring these tasks to use the same class loader. In the future
work section. Gong et al. [12] mentions separating current system classes (such as
file classes) into separate tasks, in accordance with the principle of least privilege.
The J-Kernel already moves facilities for files and networking out of the system
classes and into separate tasks.
A number of related safe-language systems are based on the idea of using
object references as capabilities. Wallach et. al. [35] describe three models of
Java security: type hiding (making use of dynamic class loading to control a
domain's namespace), stack introspection, and capabilities. They recommended
a mix of these three techniques. The E language from Electric Communities [8]
is an extension of Java targeted towards distributed systems. E's security archi-
tecture is capability based; programmers are encouraged to use object references
as the fundamental building block for protection. Odyssey [11] is a system that
supports mobile agents written in Java; agents may share Java objects directly.
Hagimont et al. [14] describe a system to support capabilities defined with spe-
cial IDL files. All three of these systems allow non-capability objects to be passed
directly between domains, and generally correspond to the share anything ap-
proach described in Section 2. They do not address the issues of revocation,
domain termination, thread protection, or resource accounting.
The SPIN project [3] allows safe Modula-3 code to be downloaded into the
operating system kernel to extend the kernel's functionality. SPIN has a par-
ticularly nice model of dynamic linking to control the namespace of different
extensions. Since it uses Modula-3 pointers directly as capabilities, the limita-
tions of the share anything approach apply to it.
Several recent software-based protection techniques do not rely on a partic-
ular high level language like Java or Modula-3. Typed assembly language [26]
pushes type safety down to the assembly language level, so that code written at
the assembly language level can be statically type checked and verified as safe.
 389

Software fault isolation [34] inserts run-time "sandboxing" checks into binary
executables to restrict the range of memory that is accessible to the code. With
suitable optimizations, sandboxed code can run nearly as fast as the original
binary on RISC architectures. However, it is not clear how to extend optimized
sandboxing techniques to CISC architectures, and sandboxing cannot enforce
protection at as fine a granularity as a type system. Proof carrying code [27, 28]
generalizes many different approaches to software protection—arbitrary binary
code can be executed as long as it comes with a proof that it is safe. While
this can potentially lead to safety without overhead, generating the proofs for a
language as complex as Java is still a research topic.
The J-Kernel enforces a structure that is similar to traditional capability sys-
tems [22, 23]. Both the J-Kernel and traditional capability systems are founded
on the notion of unforgeable capabilities. In both, capabilities name objects in a
context-independent manner, so that capabilities can be passed from one domain
to another. The main difference is that traditional capability systems used virtual
memory or specialized hardware support to implement capabilities, while the J-
Kernel uses language safety. The use of virtual memory or specialized hardware
led either to slow cross-domain calls, to high hardware costs, or to portability
limitations. Using Java as the basis for the J-Kernel simplifies many of the is-
sues that plagued traditional capability systems. First, unlike systems based on
capability lists, the J-Kernel can store capabilities in data structures, because
capabilities are implemented as Java objects. Second, rights amplification [22]
is implicit in the object-oriented nature of Java: invocations are made on meth-
ods, rather than functions, and methods automatically acquire rights to their
self parameter. In addition, selective class sharing can be used to amplify other
parameters. Although many capability systems did not support revocation, the
idea of using indirection to implement revocation goes back to Redell [31]. The
problems with resource accounting were also on the minds of implementers of
capability systems—Wulf et. al. [36] point out that "No one 'owns' an object in
the Hydra scheme of things; thus it's very hard to know to whom the cost of
maintaining it should be charged".
Single-address operating systems, like Opal [6] and Mungi [17], remove the
address space borders, allowing for cheaper and easy sharing of data between
processes. Opal and Mungi were implemented on architectures offering large
address spaces (64-bit) and used password capabilities as the protection mech-
anism. Password capabilities are protected from forgery by a combination of
encryption and sparsity.
Several research operating systems support very fast inter-process commu-
nication. Recent projects, like L4, Exokernel, and Eros, provide fine-tuned im-
plementations of selected IPC mechanisms, yielding an order of magnitude im-
provement over traditional operating systems. The systems are carefully tuned
and aggressively exploit features of the underlying hardware.
The L4 /i-kernel [15] rigorously aims for minimality and is designed from
scratch, unlike first-generation /i-kernels, which evolved from monolithic OS
kernels. The system was successful at dispelling some common misconceptions
 390

about /i-kernel performance limitations. Exokernel [9] shares L4's goal of being
an ultra-fast "minimalist" kernel, but is also concerned with untrusted loadable
modules (similar to the SPIN project). Untrusted code is given efficient control
over hardware resources by separating management from protection. The focus
of the EROS [33] project is to support orthogonal persistence and real-time com-
putations. Despite quite different objectives, all three systems manage to provide
very fast implementations of IPC with comparable performance, as shown in Ta-
ble 6. A short explanation of the 'operation' column is needed. Round-trip IPC
is the time taken for a call transferring one byte from one process to another and
returning to the caller; Exokernel's protected control transfer installs the callee's
processor context and starts execution at a specified location in the callee.

System Operation Platform Time

L4 Round-trip IPC P5-133 1.82^8
Exokernel Protected control transfer (r/t) DEC-5000 2.40/iS
Eros Round-trip IPC P5-120 4.90^s
J-Kernel Method invocation with 3 args P5-133 3.77//S
Table 6. Comparison with selected kernels.

The results are contrasted with a 3-argument method invocation in the J-

Kernel. The J-Kernel's performance is comparable with the three very fast sys-
tems. It is important to note that L4, Exokernel and Eros are implemented as
a mix of C and assembly language code, while J-Kernel consists of Java classes
without native code support. Improved implementations of JVMs and JITs are
likely to enhance the performance of the J-Kernel.

6 Conclusion
The J-Kernel project explores the use of safe language technology to construct
robust protection domains. The advantages of using language-based protection
are portability and good cross-domain performance. The most straightforward
implementation of protection in a safe language environment is to use object ref-
erences directly as capabilities. However, problems of revocation, domain termi-
nation, thread protection, and resource accounting arise when non-shared object
references are not clearly distinguished from shared capabilities. We argue that
a more structured approach is needed to solve these problems: only capabilities
can be shared, and non-capability objects are confined to single domains.
We developed the J-Kernel system, which demonstrates how the issues of
object sharing, class sharing, thread protection, and resource management can
be addressed. As far as we know, the J-Kernel is the first Java-based system
that integrates solutions to these issues into a single, coherent protection system.
Our experience using the J-Kernel to extend the Microsoft IIS web server and
to implement an extensible web and telephony server leads us to believe that a
safe language system can achieve both robustness and high performance.
 391

Because of its portability and flexibility, language-based protection is a natu-

ral choice for a variety of extensible applications and component-based systems.
From a performance point of view, safe language techniques are competitive
with fast microkernel systems, but do not yet achieve their promise of making
cross-domain calls as cheap as function calls. Implementing a stronger model
of protection than the straightforward share anything approach leads to thread
management costs and copying costs, which increase the overhead to much more
than a function call. Fortunately, there clearly is room for improvement. We
found that many small operations in Java, such as allocating an object, invoking
an interface method, and manipulating a lock were slower than necessary on
current virtual machines. Java just-in-time compiler technology is still evolving.
We expect that as virtual machine performance improves, the J-Kernel's cross-
domain performance will also improve. In the meantime, we will continue to
explore optimizations possible on top of current off-the-shelf virtual machines,
as well as to examine the performance benefits that customizing the virtual
machine could bring.

Acknowledgments Portions of this paper were originally published by the

USENIX Association in Proceedings of the USENIX 1998 Annual Technical Con-
ference, New Orleans, Louisiana, June 15-19, 1998.
The authors would like to thank Greg Morrisett, Fred Schneider, Fred Smith,
Lidong Zhou, and the anonymous reviewers for their comments and suggestions.
This research is funded by DARPA ITO Contract ONR-N00014-95-1-0977, NSF
CAREER Award CCR-9702755, a Sloan Foundation fellowship, and Intel Corp.
hardware donations. Chi-Chao Chang is supported in part by a doctoral fellow-
ship (200812/94-7) from CNPq/Brazil.

References
1. G. Back, P. Tullmann, L. StoUer, W. C. Hsieh, J. Lepreau. Java Operating Sys-
tems: Design and Implementation. Technical Report UUCS-98-015, Department of
Computer Science, University of Utah, August, 1998.
2. D. Balfanz, and Gong, L. Experience with Secure Multi-Processing in Java. Tech-
nical Report 560-97, Department of Computer Science, Princeton University,
September, 1997.
3. B. Bershad, S. Savage, P. Pardyak, E. Sirer, M. Fiuczynski, D. Becker, S. Eggers,
and C. Chambers. Extensibility, Safety and Performance in the SPIN Operating
System. 15th ACM Symposium on Operating Systems Principles, p.267-284, Cop-
per Mountain, CO, December 1995.
4. B. Bershad, T. Anderson, E. Lazowska, and H. Levy. Lightweight Remote Proce-
dure Call. 12th ACM Symposium on Operating Systems Principles, p. 102-113,
Lichtfield Park, AZ, December 1989.
5. R. S. Boyer, and Y. Yu. Automated proofs of object code for a widely used micro-
processor. J. ACM 43(1), p. 166-192, January 1996.
6. J. Chase, H. Levy, E. Lazowska, and M. Baker-Harvey. Lightweight Shared Ob-
jects in a 64-Bit Operating System. ACM Object-Oriented Programming Systems,
Languages, and AppHcations (OOPSLA), October 1992.
 392

7. G. Czajkowski and T. von Eicken. JRes: A Resource Accounting Interface for Java.
To appeax in proceedings of the 1998 Conference on Object-Oriented Programming
Languages, Systems, and Applications.
8. Electric Communities. The E White Paper.
http://www.communities.eom/products/tools/e.
9. R. Engler, M. Kaashoek, and J. James O'Toole. Exokernel: An Operating System
Architecture for Application-Level Resource Management. 15th ACM Symposium
on Operating Systems Principles, p. 251266, Copper Mountain, CO, December
1995.
10. B. Ford, G. Back, G. Benson, J. Lepreau, A. Lin, and O. Shivers. The Fluke OSKit:
A substrate for OS and language research. In Proc. Of the 16th SOSP, pp. 38-51,
St. Malo, Prance, October 1997.
11. General Magic. Odyssey, http://www.genmagic.com/agents.
12. L. Gong, and Schemers, R. Implementing Protection Domains in the Java Devel-
opment Kit 1.2. Internet Society Symposium on Network and Distributed System
Security, San Diego, CA, March 1998.
13. J. Gosling, B. Joy, and G. Steele. The Java language specification. Addison-Wesley,
1996.
14. D. Hagimont, and L. Ismail. A Protection Scheme for Mobile Agents on Java.
3rd Annual ACM/IEEE Int'l Conference on Mobile Computing and Networking,
Budapest, Hungary, September 2630, 1997.
15. H. Haertig, et. al. The Performance of fi-Kernel-Based Systems. 16th ACM Sym-
posium on Operating Systems Principles, p. 6677, Saint-Malo, FYance, October
1997.
16. C. Hawblitzel, C. C. Chang, G. Czajkowski, D. Hu, and T. von Eicken. Imple-
menting Multiple Protection Domains in Java. 1998 USENIX Annual Technical
Conference, p. 259-270, New Orleans, LA, June 1998.
17. G. Heiser, et. al. Implementation and Performance of the Mungi Single-Address-
Space Operating System. Technical Report UNSW-CSE-TR-9704, Univeristy of
New South Wales, Sydney, Austraha, June 1997.
18. JavaSoft. Java Telephony API. http://java.sun.com/products/jtapi/index.html.
19. JavaSoft. Remote Method Invocation Specification, http://java.sun.com.
20. JavaSoft. New Security Model for JDK1.2. http://java.sun.com
21. JavaSoft. Java Servlet API. http://java.sun.com.
22. A. K. Jones and W. A. Wulf Towards the Design of Secure Systems. Software
Practice and Experience, Volume 5, Number 4, p. 321336, 1975.
23. H. M. Levy. Capability-Based Computer Systems. Digital Press, Bedford, Mas-
sachusetts, 1984.
24. J. Liedtke, et. al. Achieved IPC Performance. 6th Workshop on Hot Topics in
Operating Systems, Chatham, MA, May.
25. Microsoft Corporation. Microsoft Security Management Architecture White Paper.
http://www.microsoft.com/ie/ security.
26. G. Morrisett, D. Walker, K. Crary, and N. Glew. Prom System F to Typed Assembly
Language. 25th ACM Symposium on Principles of Programming Languages. San
Diego, CA, January 1998.
27. G. Necula and P. Lee. Safe Kernel Extensions Without Run-Time Checking.
2nd USENIX Symposium on Operating Systems Design and Implementation, p.
229243, Seattle, WA, October 1996.
28. G. Necula. Proof-carrying code. 24th ACM Symposium on Principles of Program-
ming Languages, p. 106119, Paris, 1997.
 393

29. Netscape Corporation. Java Capabilities API. http://www.netscape.coni.

30. Rashid, R. Threads of a New System. Unix Review, p. 3749, August 1986.
31. D. D. Redell. Naming and Protection m Extendible Operating Systems. Technical
Report 140, Project MAC, MIT 1974.
32. Z. Shao. T'yped Common Intermediate Format. 1997 USENIX Conference on
Domain-Specific Languages, Santa Barbara, California, October 1997.
33. J. S. Shapiro, D. J. Farber, and J. M. Smith. The Measured Performance of a
Fast Local IPC. 5th Int'l Workshop on Object-Orientation in Operating Systems,
Seattle, WA. 1996
34. R. Wahbe, S. Lucco, T. E. Anderson, and S. L. Graham. Efficient Software-Based
Fault Isolation. 14th ACM Symposium on Operating Systems Principles, p. 203216,
Asheville, NC, December 1993.
35. D. S. Wallach, D. Balfanz, D. Dean, and E. W. Felten. Extensible Security Ar-
chitectures for Java. 16th ACM Symposium on Operating Systems Principles, p.
116128, Saint-Malo, France, October 1997.
36. W. A. Wulf, R. Levin, and S.P. Harbison. Hydra/C. mmp: An Experimental Com-
puter System, McGraw-Hill, New York, NY, 1981.
 Secure Network Objects*

Leendert van Doom, Martin Abadi, Mike Burrows, and Edward Wobber

Abstract. We describe the design and implementation of secure net-

work objects, which provide security for object-oriented network commu-
nication. The design takes advantage of objects and subtyping to present
a simple but expressive programming interface for security, supporting
both access control lists and capabilities. The implementation of this
design fits nicely within the structure of the existing network objects
system; we discuss its internal components, its performance, and its use
in some applications.

1 Introduction
Object-oriented communication has become popular in distributed systems [2,
23,19]. With objects or without them, distributed systems typically rely on
networks with no low-level support for security; the vulnerability of distributed
systems is by now evident and worrisome [24,4]. Therefore, a need exists for
secure object-oriented communication.
We describe the design and implementation of secure network objects. Secure
network objects extend Modula-3 network objects [18,2] with security guaran-
tees. When a client invokes a method of a secure network object over the network,
the main security properties are:
- The client must possess an unforgeable object reference.
- The client and the owner of the object can choose to authenticate each other.
- The arguments and results of the method invocation are protected against
tampering and replay, and optionally against eavesdropping.
For high-speed bulk communication, the network objects system supports buff-
ered streams called readers and writers. We make these streams secure also.
Our design accommodates both access control lists (ACLs) [11] and capa-
bilities [6]. It seems natural to treat network object references as capabilities;
moreover, these capabilities can be implemented efficiently. However, capabili-
ties suffer from the well-known confinement problem: it is hard to keep them
sufficiently secret (cf. [10]). The support for ACLs allows implementors to limit
this problem, and to use identity-based security whenever that is appropriate, in
particular for auditing. Systems with both ACLs and capabilities are not new;
we include some comparisons in section 6.
Based on "Secure Network Objects" by Leendert van Doom, Martin Abadi, Mike
Burrows, and Edward Wobber, which appeared in the Proceedings of the IEEE
Symposium on Security & Privacy; Oakland, California, May 1996; 211-221. ©1996
IEEE.
 396

The central goal of our work was the integration of security and network
objects. We have obtained the following features:
- Apphcations can use security easily, with minimal code changes.
* Security is mostly encapsulated within the network objects run-time sys-
tem.
* Objects and methods provide convenient units of protection.
* Subtyping expresses security properties quite simply. Secure network ob-
jects are a subtype of regular network objects.
- Through the combination of ACLs and capabilities, the security model is
rich enough to enable applications with sophisticated security requirements.
- The implementation of our design is reasonably straightforward. In this re-
spect, we benefited from the structure of the existing network objects system.
We also borrowed ideas from previous work on authentication [27]: each node
runs an authentication agent that is responsible for managing keys and for
identifying local users to other nodes. We feel that our experience partially
validates those previous efforts.
The next section presents our programming interface. Sections 3 and 4 de-
scribe the two main components of our system: the authentication agent and the
run-time system. Section 5 discusses experience with secure network objects, in-
cluding performance measurements and some example applications. Section 6
discusses related work.

2 Programming Interface
In a world with fast CPUs, no government export controls, and pervasive use
of cryptographic credentials, we would give uniform security guarantees for all
objects, even for those that do not require them. This would make for a simpler
system and a shorter paper.
As a compromise, we define network objects with three levels of security:
(1) no security; (2) authenticity; (3) authenticity and secrecy. We call the last
two kinds secure network objects. A secure invocation is the invocation of a
method of a secure network object.
An important characteristic of our design is that it gives security guarantees
for whole objects rather than for individual methods. This allows us to specify the
properties of secure network objects through subtyping, rather than by inventing
new language features. We have a type of objects with no security, a subtype
with authenticity, and a further subtype that adds secrecy.
These types are explained in the following sections, along with a type to
represent identities. We postpone the discussion of readers and writers to sec-
tion 2.6.

2.1 Network Objects (Background)

A Modula-3 object is a reference (pointer) to a data record paired with a method
suite. The method suite is a record of procedures that take the object as first
 397

parameter. A method invocation specifies a method name, an object, and addi-

tional arguments. It may yield results, consisting of a return value and values
for any VAR parameters.
A network object is a reference meaningful throughout a network; its methods
can be invoked from multiple address spaces, possibly on different nodes. A
method invocation is rem,ote if it crosses an address-space boundary. Method
invocations have at-most-once semantics (at least in the absence of attackers).
Each network object has one address space as owner (or server); the object
always has the same owner. Other address spaces are clients. Typically, each
host exports a well-known network object that acts as a local name server. To
import an object, a client may contact the name server of the owner of the object,
or it may receive the object as argument or result of a remote invocation.
Each client of a network object has a special object, the surrogate; passing a
network object to a new client causes a corresponding surrogate to be created at
the client. The client invokes the methods of the surrogate, which in turn invoke
the methods of the object at the owner. The presence of a surrogate is invisible
to the programmer.
An object type specifies a collection of methods. A new object type can be
defined as a subtype of an existing object type; the new type may inherit some
methods from the existing type, may override others, and may add some methods
of its own. The type NetObj .T is the base type for the network object subtyp-
ing hierarchy: all types of network objects are subtypes of NetObj .T; the type
NetObj .T does not specify any methods. In general, network objects of type
NetObj . T are not secure.

2.2 Authenticity

We introduce a subtype AuthNetObj .T of NetObj .T. Network objects of type

AuthNetObj .T are unforgeable references. When a chent invokes a method of an
object of type AuthNetObj .T, the following guarantees hold (even in the presence
of attackers):

1. Integrity: The invocation that the server receives is exactly the one issued
by the client. The results that the client receives are exactly the ones issued
by the server as a response to this invocation.
2. At-most-once semantics: The server receives the invocation at most once.
The client receives the response at most once.
3. Confinement: If the invocation or the response contains a secure network
object, an eavesdropper does not learn enough to invoke a method of that
object.

By (1), the server knows that the client has the method name, the object,
and all additional arguments of the invocation; the client knows that the server
has the results. In addition, since an object has a unique server, that same server
responds to all invocations of methods of the object.
By (2), client and server are protected against replays.
 398

By (3), a secure network object can be treated as a capability or protected

name [17]. If a secure network object is passed only from the server to a single
client in a secure invocation, the server knows that any later invocation using
this object originates in this client. More generally, if the object is passed only
in secure invocations between trusted parties, the server knows that any later
invocation using this object originates in a trusted party.
One could imagine extending (3) to insecure invocations; one would protect
secure network objects even when they are passed in insecure invocations. How-
ever, there is not much gain in doing this. For instance, the result of an insecure
invocation may be a secure network object; there is little point in protecting this
result against eavesdroppers because any address space could invoke the method
that gives this result.

2.3 Secrecy
We introduce a subtype SecNetObj .T of AuthNetObj .T for secret communica-
tion. When a client invokes a method of an object of type SecNetObj .T, the
following additional guarantee holds:
4. Secrecy: An eavesdropper does not obtain any part of the method name, the
object, the additional arguments, or the results of the invocation.
However, we do not attempt to provide perfect anonymity. An eavesdropper may
recognize that two method invocations are for the same object. As explained in
the next section, an eavesdropper may also learn who is communicating.

2.4 Identity
An identity consists of a user name and a host name. For simplicity, we assume
that each address space is running on behalf of one user on one host, and we
associate with each address space the corresponding identity.
We introduce a type to represent identities:

INTERFACE Ident;

TYPE
T <: OBJECT METHODS
userNameO : TEXT;
hostNameO : TEXT;
END;

PROCEDURE MineO: T;

END Ident.

This interface describes a type I d e n t . T, exposing the methods userNeune and

hostName. (ident .T is an object type, but not a network object type.) Each of
 399

userName and hostName takes no arguments and returns a string. An identity

object is an object of type I d e n t . T. The interface also provides a procedure
I d e n t . Mine; a call to I d e n t . Mine in an address space always returns the identity
of the address space. Attempting to pass any identity other than Ident.MineO
in a remote invocation is a run-time error that raises an exception; this behavior
is chosen for programmer convenience, but is not necessary for security.
The following guarantee is associated with the type I d e n t . T:

5. If an identity object is passed as argument or result of a secure remote

invocation, the receiver is assured that calls on userName and hostName will
return the identity of the sending address space.

In contrast, an identity object received in an insecure invocation may not identify

its sender—the sender may have been dishonest.
Even if an invocation is secure, the anonymity of an address space that sends
Ident.MineO is not protected: an eavesdropper can obtain the corresponding
user name and host name. In fact, an eavesdropper can deduce host names
simply from the pattern of communication between address spaces, even when
no identities are sent. On the other hand, if an interface does not require the
exchange of identities, its users can remain anonymous; in principle, an address
space discloses the name of its user only as a deliberate step.
Our treatment of identity suffices for our current applications and enables
us to take advantage of existing mechanisms such as Kerberos [25] or Sun's
secure RPC system [26]. However, it would be straightforward to elaborate our
notion of identity and to allow each address space to have multiple identities.
In particular, we could borrow from the work of Lampson, Wobber, et al. [12,
27]; our type Ident.T is a simplified version of the type Auth described in that
work.
It is not hard to imagine other schemes for communicating identities. For
example, identity objects could be passed explicitly or implicitly on every call.
In choosing our scheme, we have been careful to avoid language changes, and
have attempted to minimize overhead.

2.5 Discussion: Using Capabilities, Identities, and ACLs

One of the main applications of identity objects is in bootstrapping trust. A

client typically obtains its first secure network object as result of an invocation
on an insecure network object, often a name server. A priori, there is no trust
between the owner of the secure network object and the chent. This trust is
established when client and server exchange and test identity objects. Once this
has been done, they may choose to pass other secure network objects. Because
these objects are obtained as part of secure invocations, further checks of identity
may not be necessary.
For example, an insecure name server may export a secure network object
f s of type F i l e S e r v e r .T providing access to a remote file server:
 400

INTERFACE FileServer;

IMPORT AuthNetObj , Ident, Buffer;

TYPE
T = AuthNetObj.T OBJECT METHODS
owner0: Ident.T;
openCid: Ident.T; name: TEXT): File;
END;

File = AuthNetObj.T OBJECT METHODS

readO : Buffer.T;
write(Buffer.T): INTEGER;
END;

END FileServer.

By convention this object provides an owner method which returns the identity
of the principal on whose behalf the file server is running. This result is used by
a client to authenticate the file server.
Although f s is secure, anyone can obtain f s and invoke its methods. There-
fore, it is reasonable for the file server to expect an identity as argument of any
method invocation. When a client c issues the call f s . o p e n ( I d e n t . M i n e ( ) ,
" / e t c / m o t d " ) , the open method should check the identity of c. If this check
succeeds, the open method may return another secure network object f, which
represents the open file "/etc/motd". Client c may then invoke f . r e a d ( ) . Be-
cause c and the file server have authenticated one another, c and the owner of
f may choose not to authenticate one another further, even though the owner
of f need not be the file server.
Identity objects have important applications beyond bootstrapping. They
allow a server to confine the use of capabilities to particular clients. Additionally,
identities can be logged to construct an audit trail.
In our example, c could pass its identity to the read method of f, which
could check that c is a particular user, or belongs to a particular group. The
identity check provides protection even if c publishes f. The read method could
also record the names of all callers, for future inspection.
Another important application of the type Ident.T is the implementation
of various authorization mechanisms, and particularly reference monitors with
ACLs. Several designs are possible. For example, a secure network object may
include a method checkACL for making access-control decisions; the arguments
of checkACL are a mode m (such as read or write) and an identity i; the result
is a boolean that indicates whether the user with identity i is allowed access
with mode m to the object. In the intended implementation, checkACL compares
i with names kept in fists (that is, in ACLs). Additional methods enable the
modification of the ACLs. When group-membership checks are needed, checkACL
can consult group registries across the network using secure invocations.
 401

2.6 Readers and Writers

Modula-3 includes buffered streams, called readers and writers. For example, a
reader might be the stream of data from a file or from a terminal. The network
objects system allows readers and writers to be passed between address spaces,
as follows: an address space creates a reader or writer, and passes it to one other
address space, which reads from it or writes to it, but never passes it. The two
address spaces can then use the reader or writer to transmit data directly on their
underlying network connection, without the overhead of method invocations.
Readers and writers are treated specially when passed in secure invocations:

6. If a reader or writer rw is passed as argument or result of a secure invocation,

operations on rw have the same security guarantees as the invocation.
For instance, if a reader rd is passed as argument in the invocation o .m(rd), and
o is of type AuthNetObj .T, then operations on rd are secured in the same way
as operations on an object of type AuthNetObj .T. Therefore, data read from rd
is protected just as if it had been passed as an argument in the invocation. The
example given in section 2.7 further illustrates the use of readers and writers.
An alternative treatment could be based on new types for readers and writ-
ers with security properties (analogous to AuthNetObj.T and SecNetObj .T).
Modula-3 does not support multiple inheritance, so introducing these new types
could have been troublesome. We have chosen our scheme in order to allow
existing classes of readers and writers to be made secure without modification.

2.7 An Example
Let us consider a trivial terminal server that offers shells for remote users. Be-
fore a user gets a shell, the server obtains the user's identity, both to verify that
the user is legitimate and to associate the identity with the shell. On the other
hand, the user obtains the server's identity, and can check that the server is
the expected one and not an impostor. Once the user has the shell, the user's
commands and their results are protected against tampering, replay, and eaves-
dropping.
The terminal server exports the following interface:

INTERFACE STS;

IMPORT SecNetObj, Ident, Rd, Wr;

TYPE
T = SecNetObj.T OBJECT METHODS
owner0: Ident.T;
create_shell(id: Ident.T; rd: Rd.T; wr: Wr.T);
END

END STS.
 402

This interface declares the object type STS.T as a subtype of SecNetObj .T

with two new methods, owner and c r e a t e _ s h e l l . The owner method returns
the identity of the server (by calUng Ident.Mine). The c r e a t e . s h e l l method
creates a connection, starts a shell with the identity of the user, and connects
the streams rd and wr as standard input and output, respectively.
A simple client program might be:

MODULE C l i e n t EXPORTS Main;

IMPORT NetObj , I d e n t , STS, S t d i o ;

CONST serverHost = " f o o . b a r . l a d i d a " ;

VAR
agent: NetObj.Address;
s t s : STS.T;
server_ident: Ident.T;
BEGIN
agent := NetObj.Locate(serverHost);
sts := NetObj.Import("STS", agent);
server_ident := sts.ownerO;
IF server_ident.userNameO = "root"
AND server_ident.hostNameO = serverHost
THEN
sts.create_shell(Ident.Mine 0 ,
Stdio.stdin, Stdio.stdout);
ELSE
(* the server is an impostor *)
END;
END Client.
The client program imports an object s t s of type STS. T and verifies the iden-
tity s e r v e r . i d e n t of its owner. If this check succeeds, the client program starts
a shell by calling s t s . create_shell. Since STS.T is a subtype of SecNetObj .T,
the shell's standard input and output benefit from the guarantees associated
with SecNetObj .T.

3 Authentication Agents
So far we have focused on the design of a programming interface for security; in
the remainder of the paper we describe our implementation for this programming
interface. Our system has two main components, the authentication agent and
the run-time system. We describe the first in this section and the second in the
next.
In our system, each node runs an authentication agent. An authentication
agent is a process that assists application address spaces for the purposes of
 403

security. It communicates with its local clients only via local secure channels.
In our implementation, the authentication agent is a user-level process; as local
secure channels we have used Unix domain sockets in one version of our im-
plementation and System V streams in another. Each agent is responsible for
managing identities and keys, as follows.
When an address space receives an identity object, it can ask the local agent
for the corresponding user name and host name. The agent answers this question
by communicating with its peers; each agent knows the name of its host and the
user names associated with its clients.
The agent also provides channel keys. A channel key is an encryption key
shared by two address spaces. The agent performs key exchange with its peers
in order to generate these channel keys for its clients.
When the agent negotiates a new channel key, it also negotiates an expiry
time for the key and a key identifier. A key identifier allows the sender of a
message to tell the receiver which key was used for constructing the message.
The key identifier can be transmitted in clear as part of the message, while
the key itself should not be. The agent picks key identifiers so that they are
unambiguous.
Our design encapsulates all of the key exchange machinery in the authenti-
cation agent. We have tried two implementations of key exchange, one based on
our own protocol and another that relies on Sun's secure RPC authentication
service. The change of implementation was transparent to user address spaces.
In the second implementation, we wrote 1400 lines of C code for the agent.
We took the idea of using an authentication agent from the work of Wobber et
al. [27]. The authentication agent described in that work is more elaborate than
ours; for example, it deals with delegation and supports channel multiplexing.
These features could be incorporated in our system, though they may preclude
the use of off-the-shelf authentication software.

4 The Run-time System

In this section we describe the security-related code that runs in application ad-
dress spaces. We adapted the original code of Modula-3 network objects, making
a few changes:
- We defined secure network object references to be capabilities.
- We extended the protocol used by network objects with security information
and with functions like message digesting and encryption.
- We modified the run-time system, adding marshaling code and cryptography.
The changes were relatively small. We added 2500 lines of Modula-3 code to the
run-time system. We also integrated public-domain implementations of DES [16]
and MD5 [22] in C. We did not modify the overall structure of network objects.

4.1 Capabilities
The wire representation of an insecure network object is a pair (s, objid), where:
 404

- s is an address space identifier for the owner of the object,

- objid is an object identifier, which distinguishes this object from others with
the same owner.

The wire representation of a secure network object is a tuple

(s, objid, capid, key, exp)

with the following additional components:

- capid is a capability identifier,

- key is a key associated with capid,
- exp is an expiration time.

The tuple (s, objid, capid, key, exp) is a capability. The capability is created by the
owner of the object. The key is the secret that is shared between the holders of the
capability and the owner of the object. The key may never appear in clear on the
network: it is encrypted using a channel key when transmitted between address
spaces in secure invocations. The components s, objid, and capid determine key
uniquely, and thus serve as key identifier.
The capability becomes invalid once its expiration time has passed. The use
of an expiration time can be beneficial in revoking a capability; it also limits
the use of the key. The client run-time system refreshes the capabilities it holds
before they expire. This is transparent to the client application. Each secure
network object has a hidden method that the client run-time system can call to
obtain a fresh capability for the object.
In general, more than one capability may be in use for any given object. The
owner of an object maintains a set of valid capabilities for the object, and is
careful not to give out capabilities that are about to expire.

4.2 Protocol

If a client c holds a network object reference (s, objid) for an object of type
Net Ob j .T, an invocation of a method of the object consists of a request from c
and a reply from the owner s. The request is (Request: c, s, objid, reqdata), where
reqdata includes the method name and arguments of the invocation. The reply
is (Reply: c, s, repdata), where repdata contains the results of the invocation.
If a client c holds a capability (s, objid, capid, key, exp) for an object of type
AuthNetObj .T, both the request and the reply are modified. We assume that the
local authentication agents for c and s have authenticated one another, and that
they make available a channel key chankey for communication between c and s.
This key will be used for signing the request and the reply, and sometimes for
encryption, as follows.
The request has two parts, a body and a signature. The body, reqbody, is:

(Request: c, s, objid, capid, mid, reqdata)

 405

where mid is a message identifier that c has never before attached to a request
signed using chankey. The signature is:

Hash{reqbody, chankey, key)

where Hash is a cryptographic hash function (a one-way message digest function

such as MD5). Upon receipt of the request, s verifies that mid is not a duphcate
and checks the signature.
Like the request, the reply has a body and a signature. The body, repbody,
is:
(Reply: c,s, mid, repdata)
The signature is:
Hash{repbody, chankey, key)
Upon receipt of the reply, the client verifies that c, s, and mid match those of
the request, and checks the signature.
Both reqdata or repdata may contain capabilities and identity objects. These
are treated specially:
— On the wire, capability keys are encrypted under chankey.
- The wire representation of an identity object is simply a placeholder. When c
(or s) receives such a placeholder, it replaces the placeholder with an identity
object constructed locally; the methods of this identity object call the local
authentication agent to obtain the user name and the host name associated
with s (or c, respectively).
For an object of type SecNetObj .T, the protocol is the same, except that
both the request and the reply are completely encrypted under the channel key
chankey.
So far we have ignored readers and writers, because their discussion is in-
dependent of that of the protocol and because security for them is obtained by
reducing them to secure network objects. In the existing implementation of net-
work objects, passing a reader or writer rw amounts to passing a network object
called a voucher. The voucher has a type with special marshaling routines; it
mediates communication over rw. For security, we arrange that the type of the
voucher be a subtype of AuthNetObj .T or SecNetObj .T.
We can now justify the guarantees listed in section 2:
1. The signatures provide integrity. Since only c and s have chankey, only they
could have generated Hash{reqbody, chankey, key). If s never generates a re-
quest of the form (Request: c, s,...), which appears to come from c, then
s knows that only c could have generated Hash{reqbody, chankey, key), and
hence that c must have endorsed reqbody. Similarly, c knows that s must have
endorsed repbody. Moreover, the request and reply are matched because they
both include mid.
2. At-most-once semantics is guaranteed because s checks that at most one
request from c includes mid and is signed using chankey; and because c
accepts rephes only for outstanding calls.
 406

c and s demonstrate knowledge of the capability key key by transmitting

Hash(reqbody, chankey, key) and Hash{repbody, chankey,key). Thus, key does
not appear in clear on the wire. Any capability keys for other secure network
objects are protected from eavesdroppers by encryption under chankey.
Since only c and s have chankey, encryption under chankey protects the
secrecy of the invocation against eavesdroppers.
When an address space sends an identity object, the receiver constructs an
identity object that reflects the true identity of the sender. An address space
that runs our code will never pass an identity other than its own; but even if
an address space runs other code and succeeds in passing an identity other
than its own, it will not be believed.
When a reader or writer rw is passed in the invocation of a method of an
object of type AuthNetObj . T (or SecNetObj . T), the voucher for rw has type
AuthNetObj .T (or SecNetObj .T, respectively). Therefore, since operations
on rw are mediated by the voucher, they have the same security guarantees
as the invocation.

4.3 Implementation in the Run-time System

In order to implement our design, we made only a few changes to the exist-
ing network objects code. Most of these changes were in the implementation
of StubLib, the interface that the network objects run-time system provides to
stubs. (Stubs are machine-generated subroutines that are responsible for mar-
shaling and unmarshaling arguments and results of remote invocations.) Some
changes were also required in the StubLib interface itself, but they were minor;
we added only two lines of code to the stub generator to accommodate them.
The network objects system allows the use of multiple transport protocols
(such as TCP or X.25). Our security implementation is independent of the choice
of transport protocol, so one can add support for new protocols without altering
or writing security code.
In the network objects system, when two address spaces ai and a-z commu-
nicate, each of them has a connection object consisting of a reader and a writer.
Data written into ai's writer appears in 02's reader, and vice versa. We embed
the cryptography necessary for the protocol of section 4.2 in our implementa-
tion of connection objects. We do so by defining readers and writers that apply
cryptography to specified byte ranges in the data they transmit. These readers
and writers have methods for computing cryptographic digests, for encryption,
and for decryption.
As described in section 4.2, each message of our protocol includes an identifier
mid for protection against replays. Were all address spaces single threaded, a
simple sequence number would make a convenient message identifier. In order to
accommodate multi-threaded address spaces, we associate each message with a
secure channel, and let mid include a secure channel identifier. Next we explain
secure channels and their use.
At each point in time, a secure channel from an address space ai to an address
space a2 determines a channel key and a sequence number. (Two secure channels
 407

Table 1. Perforniance of secure network objects (in /us/call).

Test Old Insecure Authentic Secret

Null 708 772 870 991
Ten integer 747 830 929 1244
Existing surrogate 751 822 973 1146
New surrogate 1896 2044 2263 2430

may have the same key.) Both ai and a-z know the secure channel identifier
and keep track of the key and the sequence number associated with the secure
channel. With each method invocation from ai to a-z that uses a secure channel,
both ai and a2 increment the corresponding sequence number. When the key for
a secure channel is half way to its expiration, ai asks its authentication agent
for a new key; on seeing a new key identifier, 02 obtains the new key from its
own agent.
In the normal case, a secure invocation from ai to a2 proceeds as follows.
First ai chooses a secure channel from ai to 02 on which there is no outstanding
invocation (and sets up a new secure channel if none is available). Then ai
constructs the request using the channel key for the secure channel; mid is the
concatenation of the sequence number and the secure channel identifier. When 02
receives the request, it compares the sequence number in mid against its version
of the sequence number. The reply uses the same key and message identifier as
the request. When ai receives the reply, it verifies that the key is still current
and checks m,id.
In a more general scheme, the request and the reply may use different keys
and message identifiers. This generality has some advantages; for example, it
allows an invocation that returns after a long wait to use a fresh key. The changes
in the definition of message identifiers are considerable, so we do not discuss them
here.

5 Performance and Applications

Next, we describe our experiments with secure network objects. We discuss the
performance of our system and two example apphcations.

5.1 Performance Measurements

We measured the performance of our system with the same tests used for the
original implementation of network objects [2]. Table 1 presents a subset of our
measurements that characterizes the performance of secure network objects in
comparison with the performance of the original implementation.
We measured round-trip invocation latency for remote invocations. The first
column of Table 1 (labelled Old) gives the performance of method invocations
in the original implementation. The remaining three columns concern method
invocations in our implementation, to objects of type NetObj .T, AuthNetObj .T,
 408

and SecNetOb j . T, respectively. The rows of Table 1 indicate method invocations

with different sorts of arguments and with no results. In the null test, the method
has no arguments. In the ten integer test, the method takes ten 64-bit integers
as arguments. In the existing surrogate test, a network object is marshaled to a
receiver that already has a surrogate for it. The type of the argument matches
the security level of the invocation, so for instance in the column Authentic the
argument has type AutMetObj . T. The new surrogate test is similar, except that
the argument is unknown to the receiver; therefore, the receiver must create a
new surrogate and register it at the owner (for garbage collection).
The tests do not include the establishment of a channel key. The cost of
establishing a channel key depends on the implementation of the authentication
agent; a typical cost may be in the order of tens of milliseconds. In most cases,
this cost is insignificant when amortized over multiple secure invocations.
We performed all our measurements using DEC 3000/700 workstations, which
contain DECchip 21064a processors at 225 MHz. On these workstations, our
MD5 implementation runs at over 15 MBytes/sec; our DES implementation at
990 KBytes/sec. We used two workstations connected by an ATM network with
a point-to-point bandwidth of over 16 MBytes/sec.
Our measurements show that our system adds a non-trivial cost to insecure
invocations. This additional cost is largely due to the expanded size of the wire
representation of network objects and of packet headers, and to the associated
processing. Secure network objects require a 24-byte object representation that
identifies both the host and address space of the object owner. The packet header
format similarly requires a host and address space identifier as well as a message
identifier.
Authentic invocations add a fixed minimum MD5 overhead of nearly 40 /xs
per invocation. Each call and return packet requires at least two 64-byte MD5
calculations, and both client and server must perform these. The remainder of
the incremental cost can be attributed to the generation of a message identifier,
and the checking of timestamps, message identifiers, and digests. The calls where
an object is used as argument incur the additional overhead of encrypting each
object capability under DES.
The difference between authentic and secret calls is primarily due to DES
processing. Since the method identifier, return code, and MD5 digest are en-
crypted as well as all arguments and results, there is a minimum of 96 bytes of
DES encryption/decryption (~100 ^s) per secret invocation. (The encryption
of the MD5 digest is not necessary, and we could eliminate it for even better
performance.) The ten integer test adds 80 bytes of arguments, to be encrypted
and decrypted; this accounts for another 160 /us.
We have also measured the performance of readers and writers. Streams with
authenticity are currently limited to about 8 MBytes/sec. We observed that 40%
of the CPU cost is attributable to buffering and TCP overhead; this explains
why the bandwidth of streams with authenticity is limited to around 55% of the
raw MD5 bandwidth. Streams with both authenticity and secrecy are severely
limited by the speed of DES. According to our measurements, such streams have
 409

a throughput of about 870 KBytes/sec; this implies that nearly 90% of the CPU
is dedicated to DES computation.
As these measurements demonstrate, the performance of our system is ac-
ceptable for many applications. However, we believe that there remains much
room for optimization.

5.2 Example Applications

To date, we have had two preliminary but encouraging experiences in the appli-
cation of secure network objects.
In the first application, we have implemented a secure version of an answering
service written by Rob DeLine. This service is part of a telecollaboration system
and implements an answering machine for multi-media messages. The answering
machine is a network object with methods c r e a t e , r e t r i e v e , and d e l e t e . When
a message is created, it is stored under a special name, called a cookie; the cookie
is e-mailed to the intended recipient of the message. An e-mail reader can then
present the cookie to retrieve or to delete the message.
The original answering service is not secure. In particular, cookies are essen-
tially used as capabilities, yet they are communicated in clear and are easy to
guess. Therefore, the service has neither authenticity nor secrecy.
Our version of the answering service addresses these problems. Because of
the performance limitations of software encryption, we have not provided se-
crecy but only authenticity. In our version, the answering machine is a network
object of type AuthNetObj .T; it would be easy to make it a network object of
type SecNetObj . T instead. We store the names of both the sender and intended
recipient with each message; appropriate identity objects must be passed as ar-
guments of calls to the methods c r e a t e , r e t r i e v e , and d e l e t e . For example,
only the sender and intended recipient of a message can delete the message.
Our second application is a secure version of Obliq [3]. Roughly, Obliq is to
Modula-3 as Tel is to C. ObUq is an untyped, interpreted scripting language
that supports distributed object-oriented computation. For example, it can be
used to program computing agents that roam over a network.
In our version of Obliq, each object is implemented as a network object of type
AuthNetObj .T. When an Obhq object o is exported, it is explicitly tagged with a
list of the principals that may import it. The Obliq run-time system encapsulates
o in a reference monitor of type AuthNetObj . T. This reference monitor provides
two methods: a method for access to o and a method that returns the identity
of the owner of the reference monitor. The former method is responsible for
performing access control checks. The latter method allows a client that imports
o to verify that the server is the expected one.
We have implemented the secure version of Obliq that we have just described.
This implementation provides evidence that we can build useful security mech-
anisms for Obliq fairly easily. On the other hand, we do not yet have sufficient
experience to decide which security mechanism is most appropriate.
 410

6 Related Work

There has been much work on capabihties, in various contexts. Communication

systems with a pure capabihty model, hke that of Amoeba [15], suffer from the
confinement problem. Several restrictions and variations of the capability model
have been proposed. For example. Gong suggests adding identities to capabili-
ties [7]; Bacon et al. suggest restricting their lifetime [1]. Karger's dissertation
describes several others [9]. The ideas in our use of capabilities can be traced
back through a vast literature.
There has also been substantial work in the area of security and network
communication systems (e.g., [25,14,21]). However, to our knowledge, there is
at present no object-oriented network communication system with support for
security. Next we discuss two recent efforts in this area.
The Object Management Group has requested technology for integrating
security in CORBA [19] and has received submissions [20]; security work is cur-
rently in progress in the Object Management Group. A general architecture for
security in CORBA has also been proposed, and an implementation of the ar-
chitecture was announced as future work [5]. In summary, there seems to be
significant interest and activity on security in CORBA, but (at the time of the
writing of this paper) not yet any resolution or experimental results.
The Spring object-oriented operating system [13] supports both ACLs and
capabilities. An accurate comparison with secure network objects is difficult, as
there has not been a complete description of security in Spring. Spring has more
expressive capabilities than our system; for example, several capabilities for an
object can give different privileges. Perhaps for this reason, the combination
of identities and capabilities is less important in Spring, and has been explored
less [8]. To date, Spring security has been implemented only on a single processor;
all network communication is unprotected and capabilities are passed in clear.

7 Conclusions

Secure network objects behave like insecure network objects, but preserve the
intended semantics even in the presence of an active attacker; optionally, secure
network objects provide secrecy from eavesdroppers. In addition, identity ob-
jects form the basis for authentication in our system. The combination of secure
network objects with identity objects leads to a simple programming model and
a simple implementation, both in the spirit of the original network objects. Thus
we have integrated security and network objects.
Overall, we felt that object-orientation was helpful. Not surprisingly, we
found that objects give rise to natural units of protection. We also took ad-
vantage of the object type system to specify security properties, directly and
economically.
In our description, we have tried to be precise, but not formal. We believe that
a more formal study would be interesting. In particular, it would be worthwhile
to give notations and rules for reasoning about secure network objects.
 411

Acknowledgments

We would like to t h a n k Andrew Birrell, Greg Nelson, and Luca Cardelli for many
helpful discussions; and Roger Needham, Cynthia Hibbard, T i m M a n n , Rustan
Leino, Andrew Tanenbaum, Philip Homburg, Raoul Bhoedjang, Tim Ruhl, and
Greg Sharp for helpful comments on the paper.

References
1. Jean Bacon, Richaid Hayton, Sai Lai Lo, and Ken Moody. Extensible access control
for a hierarchy of servers. ACM Operating Systems Review, 28(3):4-15, July 1994.
2. Andrew Birrell, Greg Nelson, Susan Owicki, and Edward Wobber. Network objects.
Software Practice and Experience, S4(25):87-130, December 1995.
3. Luca Cardelli. A language with distributed scope. Computing Systems, 8{l):27-59,
January 1995.
4. W.R. Cheswick. An evening with Berferd, in which a hacker is lured, endured, and
studied. In Proceedings of the Usenix Winter '92 Conference, 1992.
5. R.H. Deng, S.K. Bhonsle, W. Wang, and A.A. Lazar. Integrating security in
CORBA based object architectures. In Proceedings of the 1995 IEEE Symposium
on Security and Privacy, pages 50-61, May 1995.
6. J.B. Dennis and E.G. van Horn. Programming semantics for multiprogrammed
computation. Communications of the ACM, 9(3):143-155, March 1966.
7. Li Gong. A secure identity-based capability system. In Proceedings of the 1989
IEEE Symposium on Security and Privacy, pages 56-63, May 1989.
8. Graham Hamilton. Personal communication, 1994 and 1996.
9. Paul Ashley Karger. Improving Security and Performance for Capability Systems.
PhD thesis, Cambridge University, October 1988.
10. Butler Lampson. A note on the confinement problem. Communications of the
ACM, 16(10):613-615, October 1973.
11. Butler Lampson. Protection. ACM Operating Systems Review, 1(8): 18-24, January
1974.
12. Butler Lampson, Martin Abadi, Mike Burrows, and Edward Wobber. Authentica-
tion in distributed systems: Theory and practice. ACM Transactions on Computer
Systems, 10(4):265-310, November 1992.
13. J. Mitchell, J. Gibbons, G. Hamilton, P. Kessler, Y. Khahdi, P. Kougiouris,
P. Madany, M. Nelson, M. Powell, and S. Radia. An overview of the Spring system.
In IEEE Compcon Spring 1994, February 1994.
14. R. Molva, G. Tsudik, E. van Herreweghen, and S. Zatti. Kryptoknight authenti-
cation and key distribution system. In Proceedings of the European Symposium on
Research m Computer Security, November 1992.
15. Sape J. Mullender, Andrew S. Tanenbaum, and Robbert van Renesse. Using sparse
capabilities in a distributed operating system. In Proceedings of the 6th IEEE
conference on Distributed Computing Systems, June 1986.
16. National Bureau of Standards. Data encryption standard. PIPS 47, 1977.
17. Roger Needham. Names. In Sape Mullender, editor. Distributed Systems, chap-
ter 12, pages 315-327. Addison-Wesley, second edition, 1993.
18. Greg Nelson, editor. Systems Programming with Modula-3. Prentice Hall, 1991.
19. Object Management Group. Common object request broker architecture and spec-
ification. OMG Document number 91.12.1.
 412

20. Object Management Group. OMG documents. See URL: http://www.omg.org/.

21. Open Software Foundation. Introduction to OSF DCE. Revision 1.0, 1992.
22. R.L. Rivest and S. Dusse. RFC 1321: The MD5 message-digest function. Internet
Activities Board, 1992.
23. Marc Shapiro. Structure and encapsulation in distributed systems: The proxy
principle. In IEEE International Conference on Distributed Computer Systems,
May 1986.
24. Eugene H. Spafford. The Internet worm program: An analysis. Computer Com-
munication Review, 19(l):17-57, January 1989.
25. J.G. Steiner, C. Neuman, and J.I. Schiller. Kerberos: An authentication service for
open network systems. In Usenix 1987 Winter Conference, pages 191-202, January
1988.
26. Sun Microsystems. RFC 1057: RFC: Remote procedure call protocol specification:
Version 2. Internet Activities Board, June 1988.
27. Edward Wobber, Martin Abadi, Michael Burrows, and Butler Lampson. Authen-
tication in the Taos operating system. ACM Transactions on Computer Systems,
12(l):3-32, February 1994.
History-Based Access Control for Mobile Code*

Guy Edjlali, Anurag Acharya, and Vipin Chaudhary

Abstract. In this chapter, we present a history-based access-control

mechanism that is suitable for mediating accesses from mobile code. The
key idea behind history-based access-control is to maintain a selective
history of the access requests made by individual programs and to use
this history to improve the differentiation between safe and potentially
dangerous requests. What a program is allowed to do depends on its own
behavior and identity in addition to currently used discriminators like
the location it was loaded from or the identity of its author/provider.
History-based access-control has the potential to significantly expand
the set of programs that can be executed without compromising security
or ease of use. We describe the design and implementation of Deeds, a
history-based access-control mechanism for Java. Access-control policies
for Deeds are written in Java, and can be updated while the programs
whose accesses are being mediated are still executing.

1 Introduction
The integration of mobile code with web browsing creates an access-control
dilemma. On one hand, it creates a social expectation that mobile code should
be as easy to download and execute as fetching and viewing a web page. On the
other hand, the popularity and ubiquity of mobile code increases the likelihood
that malicious programs will mingle with benign ones.
To reassure users about the safety of their data and to keep the user interface
simple and non-intrusive, systems supporting mobile code have chosen to err on
the side of conservatism and simplicity. Depending on its source, mobile code is
partitioned into trusted and untrusted code. Code is considered trusted if it is
loaded from disk [9,12] or if it is signed by an author/organization deemed trust-
worthy by the user [12,30]. Untrusted code is confined to a severely restricted
execution environment [9] (eg, it cannot open local files or sockets, cannot cre-
ate a subprocess, cannot initiate print requests etc); trusted code is either given
access to all available resources [30] or is given selective access based on user-
specified access-control hsts [12].
For the programs considered untrusted, these mechanisms can be overly re-
strictive. Many useful and safe programs, such as a well-behaved editor applet
from a lesser-known software company, cannot be used since it cannot open local
files. In addition, to implement new resource-sharing models such as global com-
puting [6] all communication has to be routed through brokers. This significantly

* This paper is a reprint of a paper that appeared in the Fifth ACM Conference on
Computer and Communications Security (November 3-5, 1998.)
 414

limits the set of problems that can be efficiently handled by such models. For
programs considered trusted, these models can be too lax. Errors, not just malice
aforethought, can wipe out or leak important data. Combined with a suitable
audit trail, signed programs [12] do provide the ability to take legal recourse if
need be.
In this chapter, we present a history-based access-control mechanism that is
suitable for mediating accesses from mobile code. The key idea behind history-
based access-control is to maintain a selective history of the access requests made
by individual programs and to use this history to improve the differentiation be-
tween safe and potentially dangerous requests. What a program is allowed to do
depends on its own identity and behavior in addition to currently used discrimi-
nators like the location it was loaded from or the identity of its author/provider.
History-based access-control has the potential to significantly expand the set of
programs that can be executed without compromising security or ease of use.
For example, consider an access-control policy that allows a program to open
local files for reading as long as it has not opened a socket and allows it to open
a socket as long as it has not opened a local file for reading. Irrespective of the
source of the program, such a policy can ensure that no disk-resident data will
be leaked. Strictly speaking, this is true iff it is possible to intercept all access
requests being made on behalf of the program - the requests made by itself as
well as the requests made on its behalf. The technique we present is able to
intercept all requests.
We first present some examples of history-based access-control policies. Next,
we discuss issues that have to be resolved for implementing history-based access-
control mechanisms. In section 3, we describe Deeds,^ an implementation of
history-based access-control for Java programs. Access-control policies for Deeds
are written in Java, and can be installed, removed or modified while the programs
whose accesses are being mediated are still executing. Deeds requires policies to
adhere to several constraints. These constraints are checked either at compile-
time by the Java compiler or at runtime by the Deeds policy manager. We
illustrate the operation of the Deeds user interface using snapshots. In section 4.4,
we examine the additional overhead imposed by Deeds using micro-benchmarks
as well as real programs. History-based access-control is not specific to Java or
to mobile code. It can be used for any system that allows interposition of code
between untrusted programs and protected resources. In section 5, we discuss
how a system similar to Deeds can be used to mediate accesses to OS resources
from native binaries. We conclude with a description of related work and the
directions in which we plan to extend this effort.

2 Examples

One-out-of-k: Consider the situation when you want to allow only those pro-
grams that fall into well-marked equivalence classes based on their functionality
Your deeds determine your destiny :)
 415

and behavior. For example, you want to allow only programs that provide just
the functionality of a browser or an editor or a shell. A browser can connect to
remote sites, create temporary local files in a user-specified directory, read files
that it has created and display them to the user. An editor can create local files
in user-specified directories, read/modify files that it has created, and interact
with the user. It is not allowed to open sockets. A shell can interact with the
user and can create sub-processes. It cannot open local files, or connect to remote
sites. This restriction can be enforced by a history-based access-control policy
that:

- allows a program to connect to a remote site if and only if it has neither

tried to open a local file that it has not created, nor tried to modify a file it
has created, nor tried to create a sub-process;
- allows a program to open local files in user-specified directories for modifi-
cation if and only if it has created them, and it has neither tried to connect
to a remote site nor tried to create a sub-process.
- allows a program to create sub-processes if and only if it has neither tried
to connect to a remote site nor tried to open a local file.

In effect, each program is dynamically classified into one of three equivalence

classes (browser-like, editor-like or shell-like) based on the sequence of requests
it makes. Once a program is placed in a class, it is allowed to access only the
resources that are permitted to programs in that class.

Keeping out rogues: Consider the situation where you want to ensure that
a program that you once killed due to inappropriate behavior is not allowed to
execute on your machine. This restriction can be enforced, to some extent, by
a history-based access-control policy that keeps track of previous termination
events and the identity of the programs that were terminated.

Frustrating peepers: Consider the situation where you want to allow a pro-
gram to access only one of two relations in a database but not both. One might
wish to do this if accessing both the relations may allow a program to extract
information that it cannot get from a single relation. For example, one might
wish to allow programs to access either a relation that contains the date and the
name of medical procedures performed in a hospital or a relation that contains
the names of patients and the date they last came in. Individually, these rela-
tions do not allow a program to deduce information about treatment histories of
individual patients. If, however, a program could access both relations, it could
combine the relations to acquire (partial) information about treatment histories
for individual patients. This example can be seen as an instance of the Chinese
Wall Policy [4]. To block the possibility of a hostile site being able to deduce
the same information from data provided by two different programs it provides,
programs that have opened a socket are, thereafter, not allowed to access sensi-
 416

tive relations and programs that have accessed one of the sensitive relations are,
thereafter, not allowed to open sockets.
Slowing down hogs: Consider the situation where you want to limit the rate
at which a program connects to its home site. One might wish to do this, for
example, to eliminate a form of denial of service where a program repeatedly
connects to its home site without doing anything else. This can be enforced by
a history-based access-control policy that keeps track of the timestamp of the
last request. It allows only those requests that occur after a threshold period.

3 Issues for history-based access-control

Identity of programs: Associating a content-based, hard-to-spoof identity

with a program is a key aspect of history-based access-control. That is, given
any program, it should be hard to design a substitute program for whom the
identity computation generates the same result. An important point to note is
that the code for a mobile program can come from multiple sources (from local
disk, from different servers on the network, etc). The identity mechanism should
associate a single identity with all the code that is used by a program. This is
important to ensure that a malicious program cannot assume the identity of
another program by copying parts or all of the program being spoofed.
Efficient maintenance of request-histories: Wallach et al [40] mention that
a collection of commonly used Java workloads require roughly 30000 crossings
between protection domains per CPU-second of execution. Given this request-
frequency, it is imperative that access-control checks on individual requests be
fast. Simple logging-based techniques are likely to be too expensive. Fortunately,
the request-history for many useful policies can be summarized. For example, the
request-history for a policy that allows a program to open local files for reading
if it has not opened a socket and allows it to open a socket if it has not opened a
local file for reading can be summarized by a pair of booleans - one that records
if the program has ever opened a socket and the other that records if it has ever
opened a local file.
Persistence of policies and histories: Persistent request-histories are re-
quired to block attacks that consist of running a sequence of programs each of
which makes requests that are allowed by the access-control policy but when
taken as a whole, the complete sequence of requests violates the constraints the
policy tries to enforce.
Grouping privileges: History-based mechanisms can provide extremely fine-
grain access-control. Not only is it possible to control accesses to individual
objects/resources, it is possible to differentiate between different patterns of
accesses. While this allows us to expand the set of programs that can be executed
safely, this level of flexibility can be hard to deal with. Requiring users to specify
their preferences at this level of detail is likely to be considered intrusive and
 417

therefore ignored or avoided. This problem can alleviated to some extent by

grouping acceptable patterns of program behavior and assign intuitive names to
these patterns. For example, a policy that allows a program to open no sockets,
open local files for reading, to create local files in a user-specified directory and
to open a local file for modification only if it has been been created by itself.
This set of restrictions allow a simple editor to be executed and can, jointly, be
referred to as the editor policy.

Composition and fail-safe defaults: History-based access-control policies en-

code acceptable patterns of program behavior. Different classes of programs
might have different behaviors all of which are acceptable to the user. It is,
therefore, important to provide automatic composition of multiple policies. An
important point to note here is that, by default, the access-control mechanism
should be fail-safe [32] - potentially dangerous accesses should be denied unless
explicitly granted.

4 Deeds: a history-based security manager for Java

In this section we describe the design and implementation of Deeds, a history-

based access-control mechanism for Java. We first describe the architecture of
Deeds. Next, we describe its current implementation and its user interface. In
section 4.4, we examine the performance of the mechanisms provided by Deeds.

4.1 Architecture

In this subsection, we describe the architecture of Deeds. We focus on the central

concepts of the Deeds architecture: secure program identity and security events.

Program identity The Deeds notion of the identity of a program is based on

all the downloaded code reachable during its execution. To achieve this. Deeds
performs static linking on downloaded programs, fetching all non-local code that
might be referenced. Local libraries that are part of the language implementation
(e.g Java, lang for Java, l i b c for C) are linked in as shared libraries; a separate
copy of non-system-library code is downloaded for every application that uses
it.
Deeds concatenates all non-system-library code for a downloaded program
and uses the SHA-1 algorithm [36] to compute a name for it. SHA-1 belongs
to the group of algorithms known as secure hash functions [31,36] which take
an arbitrary sequence of bytes as input and generate a (relatively) short digest
(160-bits for SHA-1). These functions are considered secure because it is compu-
tationally hard to construct two byte-sequences which produce the same digest.
In addition, the requirement, in this case, that the byte-sequences being com-
pared should represent valid programs increases the difficulty of constructing a
malicious program with the same name as a benign one.
 418

In addition to allowing a secure hash to be computed, static linking of down-

loaded code has other advantages. First, having all the code available allows
Just-in-Time compilers to perform better analysis and generate better code. Sec-
ond, it removes potential covert channels which occur due to dynamic linking -
the pattern of link requests can be used to pass information from a downloaded
program to the server(s) that is (are) contacted for the code to be linked.
Note that Java allows programs to dynamically load classes. For such pro-
grams, it is not possible, in general, to statically determine the set of classes
that might be referenced during the execution of the program. Deeds rejects
such programs and does not allow them to execute.

Security events and handlers A Deeds security event occurs whenever a

request is made to a protected resource. Examples of security events include
request to open a socket, request to open a file for reading, request to create
a file, request to open a file for modification etc. The set of security events
in Deeds is not fixed. In particular, it is not limited to requests for operating-
system resources. Programmers can associate a security event with any request
they wish to keep track of or protect.
Handlers can be associated with security events. Handlers perform two tasks:
they maintain an event-history and check whether it satisfies one or more user-
specified constraints. If any of the constraint fails, the handler raises a security-
related exception. For example, a handler for the security event associated with
opening a socket can record whether the program currently being executed has
ever opened a socket. Similarly, a handler for the security event associated with
opening a file for reading can record whether the program has opened a file for
reading.
Multiple handlers can be associated with each event. Handlers maintain sep-
arate event-histories. The checks they perform are, in effect, composed using a
"consensus voting rule" - that is, one negative vote can veto a decision and at
least one positive vote is needed to approve. In this context, a request is per-
mitted to continue if and only if at least one handler is present and none of the
handlers raises an exception.
Access-control policies consist of one or more handlers grouped together.
The handlers belonging to a single policy maintain a common history and check
common constraints. For example, the editor policy mentioned earlier would
consist of four handlers:

— a handler for socket-creation that records if a socket was ever created by this
program. It rejects the request if a file has been opened by this program (for
reading or writing).
— a handler for file-creation that associates a creator with each file created by a
downloaded program. If the file is to be created in a directory that is included
in a list of user-specified directories, it allows the request to proceed. Else,
it rejects the request.
 419

— a handler for open-file-for-read that records if a file was ever opened for
reading by this program. It rejects the request if a socket has been created
by this program.
— a handler for open-file-for-modification that records if a file was ever opened
for writing by this program. It rejects the request if a socket has been created
by this program or if the file in question was not created by this program.
Deeds allows multiple access-control policies to be simultaneously active. Poli-
cies can be installed, removed, or modified during execution. A policy is added
by attaching its constituent handlers to the corresponding events. For example,
the editor policy would be added by attaching its handlers respectively to the
socket-creation event, the file-creation event, the open-file-for-read event and the
open-file-for-modification event. Policies can be removed in an analogous manner
by detaching the constituents handlers from the associated events.
Deeds allows policies to be parameterized. For example, a policy that controls
file creation can be parameterized by the directory within which file creation is
allowed. Policies that are already installed can be modified by changing their
parameters. This allows users to make on-the-fly changes to the environment
within which mobile code executes.
Deeds provides a fail-safe default [32] for every security event. Unless overrid-
den, the default handler for an event disallows all requests associated with that
event from downloaded programs. The default handler can only be overridden
by explicit user request - either by a dialog box or by a profile file containing a
list of user preferences.

4.2 Implementation
In this subsection, we describe the implementation of Deeds. We focus on imple-
mentation of program identity, events, event-histories, policies (including con-
ventions for writing them), and policy management.

Program identity We have implemented a new class-loader for downloading

Java programs. A new instance of this class-loader is created for every down-
loaded program and is used to maintain information regarding its identity. This
class-loader statically links a downloaded program by scanning its bytecode and
extracting the set of classes that may be referred to during its execution. If the
entire program is provided as a single j a r file, this is straightforward. Else, the
class-loader fetches and analyzes the non-local classes referred to and repeats
this till transitive closure is achieved. If the scan of the bytecode indicates that
the program exphcitly loads classes, the linking operation is terminated and the
program is not allowed to run.
After the linking operation is completed, the class-loader concatenates the
code for all the non-system-library classes that are referenced by the program and
uses the implementation of the SHA-1 algorithm provided in the Java, s e c u r i t y
package to compute a secure hash. The result is used as the name of the program
and is stored in the class-loader instance created for this program. This name
 420

can be used to maintain program-specific event-histories. It can also be stored

in persistent storage and loaded as a part of the startup procedure.
Events: Two concerns guided our implementation of events: (1) the number of
events is not fixed, and (2) the number of handlers associated with individual
events could be large. We considered three alternatives. First, we could use a
general-purpose mechanism (similar to Java Beans [8] and X [33]) to register
events and handlers. The advantage of this approach is that it uses common
code to manage all events and their associated handlers; the disadvantage is
that all handlers must have the same type-signature which usually implies that
the parameters need to be packed by the event manager and unpacked by each
handler.
Second, we could dynamically modify the bytecode of the downloaded pro-
gram to insert/delete calls to handlers at the points where the events are gen-
erated (as dynamic instrumentation programs [5,15] do). To allow a user to
modify an executing policy would require us to update the bytecode of run-
ning programs. We believe that the complexity of such an implementation is not
commensurate with its advantages.
Finally, we could require that the handlers for each event be managed by
a different event manager. This approach allows us to avoid the packing and
unpacking of parameters as each event manager is aware of the parameters cor-
responding to its event. The disadvantage of this scheme is that a separate event
manager has to be written for each event. However, event managers are highly
stylized and can be automatically generated given a description of the event (see
Figure 1 for an example).
We selected the third approach for implementing security events in Deeds.
Combined with automatic generation of event managers, this allowed us to bal-
ance the needs of efficiency, implementation simplicity and ease of programming.
Examples of Deeds security events include checkReadO and checkConnectO.^
History: Given the concern about the size of the log, we have chosen to avoid
logging as a general technique for maintaining event-histories. Instead, we have
left the decision about how to store histories to individual policies. All policies
that we considered were able to summarize event-histories using simple data
structures such as counters, booleans or lists. Note that this decision has the
potential disadvantage that if policies do desire/need to log events, the lack of
a common logging mechanism can result in the maintenance of duplicate logs.
This can be fixed by using a selective logging mechanism that logs an event only
if requested to do so by one or more handlers associated with the event.

Access-control policies A Deeds access-control policy consists of the data-

structures to maintain event-histories, handlers for each of the events that are
mediated by the policy, and auxiliary variables and operations to facilitate man-
agement of multiple policies. Concretely, an access-control policy is implemented
Readers familiar with Java will recognize that all the check* () methods are security
events.
 421

public class checkReadManager implements EventManager {.

private static HandlerCheckRead hdlr = new HandlerCheckReadO ;

public static void checkRead(FileDescriptor fd.DClassLoader cl)

throws GeneralSecurityException {
for (int i=0;i<hdlr.size;i++)
hdlr.policy(i).checkRead(fd,cl);
}

public static void checkRead(String file.DClassLoader cl)

throws GeneralSecurityException {
for (int i=0;i<hdlr.size;i++)
hdlr.policy(i).checkRead(file.cl);
}

public static void checkRead(String file,Object context,

DClassLoader cl)
throws GeneralSecurityException {
for (int i=0;i<hdlr.size;i++)
hdlr.policy(i).checkRead(file,context,cl);
}
}

Fig. 1. Example of an event manager class. Managers for other events would share the
same structure but would replace checkRead by name the of the particular event. Some
administrative details have been left out of the example; these details axe common to
all event managers.

as a Java class that extends the AccessPolicy class shown in Figure 2. Handlers
are implemented as methods of this class and the event-history is implemented
as variables of this class. For example, a handler for the open-file-for-reading
event could check if a socket has yet been created by the program. If so, it could
raise a GeneralSecurityException; else, it could set a boolean to indicate that
a file has been opened for reading and return.
When a security event occurs (e.g., when checkRead is called), control is
transferred to the Deeds Security Manager which determines the class-loader for
the program that caused the event using the currentClassLoaderO method
provided by the Java Security Manager. This method returns the class-loader
corresponding to the most recent occurrence on the stack of a method from
a class loaded using a class-loader. Since a new instance of the class-loader is
created for every downloaded program and since this instance loads all non-
system-library classes for the program, currentClassLoaderO always returns
the same class-loader every time it is called during the execution of a program.
This technique safely determines the identity of the program that caused the
security event.
 422

abstract synchronized public class AccessPolicy i

public String name // name of policy instance
public Vector parameters // policy parameters
public Vector targetEvents //
public String srcFileName // source file location
// these functions have to be provided by every policy
public abstract String documentation();
public abstract void saveHistoryToDiskO ;
public abstract void restoreHistoryFromDiskO ;

public Policy(String name) {

Fig. 2. Skeleton of the AccessPolicy class. The synchronized keyword ensures that at
most one handler is updating the event-history at any given time.

Once the class-loader corresponding to the currently executing program has

been determined, the Deeds Security Manager invokes the event manager cor-
responding to event being processed (e.g., checkReadManager in Figure 1). The
event manager maintains the set of handlers associated with the event and in-
vokes the handlers in the order they v^^ere attached to the event. If any of the
handlers throws an exception, it is propagated to the caller; the remaining han-
dlers are not invoked.
Deeds policies are expected to adhere to several constraints. These constraints
are checked either at compile-time by the Java compiler or at runtime by the
Deeds policy manager. These constraints are:

— Handler methods must include a throws GeneralSecurityException clause

and must have the same name as the security event that they are intended
for. The type signature for a handler method must be the same as the type
signature of the security event it handles except for one additional argument
- the class-loader. See Figure 1 for an illustration.
— A handler method must have the same number of variants as the security
event that it is intended for. For example, a checkRead event has three vari-
ants - checkRead(FileDescriptor f d ) , checkRead(String f i l e ) , and
checkRead ( S t r i n g f i l e . Object c o n t e x t ) . Handlers for this event must
have three variants. See Figure 1 for an illustration.
— Parameters of a policy must be explicitly identified. Each parameter must
have a default value and a documentation string.
— The vector t a r g e t E v e n t s specifies which events the handlers in this policy
are to be attached to. The specification is in the form of a regular expres-
sion which is matched against fully-qualified names. For example, the target
events for a checkRead handler could be specified as ' ' F i l e l O . checkRead''
 423

or ' ' * . c h e c k R e a d ' ' . The former expression specifies only the checkRead
event defined in the FilelO package whereas the latter specifies all check-
Read events irrespective of the package they have been defined in. This spec-
ification is needed as Java's hierarchical namespace allows multiple methods
with the same name to exist in different regions of the namespace. Since a
security event is implemented by a method in a subclass of EventManager
and since every package can have its own security events, the possibility of
name clashes is real. For example, a library to perform file I/O, and a li-
brary to interact with a database could both wish to create a checkRead
event. Since packages are independently developed, extensible systems, such
as Deeds, cannot assume uniqueness of event names.
— Each policy must be accompanied by its source code and the name of the file
containing the source code should be available as a member of the class im-
plementing the policy. We believe that availability of source code of a policy
is important to instill confidence in its operation and its documentation.

Policy manager The Deeds policy manager makes extensive use of the Java re-
flection mechanism [17]. This mechanism allows Java code to inspect and browse
the structure of other classes. The policy manager uses reflection to: (1) iden-
tify methods that are to be used as handlers (they are declared p u b l i c void
and throw the GeneralSecurityException); (2) identify parameters and their
types; (3) initialize and update parameters; and (4) extract specification of the
events that the handlers are to be attached to. In addition, it performs sev-
eral administrative checks such as ensuring that all policy instances have unique
names.
The policy manager is also responsible for ensuring that policies are per-
sistent. It achieves this by storing the parameters for each policy instance on
stable storage and using them to re-install the policy when the environment is
reinitialized (on startup). It also periodically saves the event-history on stable
storage.^

4.3 User interface

The Deeds user interface comes up only on user request and is used for infre-
quent operations such as browsing/loading/installing policies. In this section, we
describe the functionality of the user interface and present snapshots.
Browsing/viewing/loading policies: The Deeds user interface allows users
to browse the set of available policies, to view documentation and source code
for these policies and to create instances of individual policies. Note that every
policy is required to have documentation (via the documentationO method)
and access to its own source code (via the srcFileName member). In addition,
every parameter has associated documentation which can be viewed. To load a
Note that individual policies are free to save the event-history as frequently as they
wish.
 424

parameterized policy, users need t o specify values for all t h e parameters of t h e

policy. Note t h a t every parameter has a default value which is displayed. T h e
Deeds policy manager uses t h e Java reflection mechanism t o figure out t h e type
of t h e parameters for display and parsing purposes. These functions of t h e user
interface a r e illustrated in figures 3 a n d 4. In Figure 3, a policy is selected by
clicking on its n a m e and operations are selected using t h e b u t t o n s . Browsing and
loading of individual pohcies is illustrated in Figure 4.

Losided Policies
Close Load Policy iBrowser
i s i i B l l Script
Save Settings Unload Policy
iFBI^clcdoor
View/Modify loaded PoMcyJLflaitage Controi
Install Policy |
Instatted Policies
View/Modify Installed Policy [editor
Uninstall Policy |

Fig. 3. Graphical interface to the Deeds Security Manager

Leakage Control Policy

Leakage Control Policy

Apply i Done Source Code Documentation |
Name ,Foo Leal<age Control
Allowed Read/Write ytmp/* Parameter Information \

Allowed Connection I*.foo.edu *.b3r,edu Parameter Information i

Fig. 4. Graphical interface for loading a policy

Installing/uninstalling policies: The Deeds user interface allows users to in-

stall loaded policies as well as to remove currently installed policies. For an
illustration, see Figure 3. A loaded policy can be installed using the I n s t a l l
 425

Policy button, and an installed policy can be removed using the U n i n s t a l l

Policy button.
Checkpointing event-histories: Deeds allows user to checkpoint the current
state of the event-histories for all policies using the Save S e t t i n g s button (see
Figure 3).
Browsing/setting default handlers: Deeds provides a fail-safe default for
every security event. Unless overridden, the default handler for an event disallows
all requests associated with that event from downloaded programs. The default
handler can only be overridden by explicit user request - either by a dialog box
or by a profile file containing a list of user preferences. The Deeds user interface
allows users to browse and set default handlers for all security events.

4.4 Performance evaluation

There are two ways in which Deeds can impact the performance of downloaded
code whose accesses it mediates. First, it can add an overhead to each request
for protected resources. Second, it can increase the startup latency as it requires
fetching, loading, linking and hashing of all non-system-library code before the
program can start executing. In this section, we evaluate the performance im-
pact of Deeds. All experiments were performed on a Sun E3000 with 266MHz
UltraSparc processors and 512MB memory.
To determine the overhead of executing Deeds security-event handlers, we
used a microbenchmark which repeatedly opened and closed files. A security
event was triggered on the request to open a file. We varied the number of han-
dlers from zero to ten. Each handler was identical (but distinct) and implemented
the e d i t o r policy described earlier in this paper. It maintains two booleans, one
that tracks if the program has ever tried to create a socket and the other that
tracks if the program has ever tried to open a file. Each time it is invoked, it
updates the file-opened boolean and checks the socket-opened boolean.
Table 1 presents the results. It shows that even with ten handlers, the over-
head of Deeds security event handlers is less than 5%. Another point to note
is that without any handlers, that is, when the infrastructure added to support
security event handlers is not used, the overhead is less than 1%.

Number of handlers 0 1 2 3 4 5 6 7 8 9 10
Percent overhead 0.7 1.8 2.6 2.4 2.9 3.5 4.2 3.9 4.3 4.1 4.3

Table 1. Overhead of Deeds security event handlers. The overhead was measured using
a microbenchmaxk which repeatedly opened and closed files. Each handler was identical
(but distinct) and implemented the editor policy.

To evaluate the impact on startup latency, we compared the time it takes to

load, analyze and link complete Java applications using the Deeds class-loader
 426

to the time it takes to load just the first file using existing class-loaders. In both
cases, all the files were local and were in the operating-system file-cache. For this
experiment, we selected seven complete Java applications available on the web.
The applications we used were: (1) news-server, the Spaniel News Server [38]
which manages and serves newsgroups local to an organization; (2) j l e x , the
JLex [23] lexical analyzer; (3) dbase, the Jeevan [22] platform-independent,
object-oriented database; (4) jawavedit, the JaWavedit audio file editor [21]
with multi-lingual voice synthesis, signal processing, and a graphical user inter-
face; (5) obf u s c a t o r , the Hashjava [14] obfuscator for Java class files; (6) javacc,
the JavaCC [20] parser generator; and (7) e d i t o r , the WingDis editor [42].
Table 2 presents results for the latency experiments. As expected, the addi-
tional startup latency increases with the number of files as well as the total size
of the program. Note this does not represent an increase in end-to-end execution
time. Existing class-loaders already parse the bytecodes of class files as a part
of the Java verification process; signed applets require computation of a simi-
lar hash function. Instead, the increase in startup latency is caused by moving
the processing for all the class files before the execution begins. We expect that,
once downloaded, programs of this size and these types (lexer/parser generators,
editors, news server, database etc) will be reused several times. In that case, the
program can be cached as a whole (instead of individual files) and the additional
startup latency has to be incurred only once.

Application newsserver j l e x dbase jawavedit obfuscator javacc e d i t o r

Number of classes 24 40 104 125 144 81 212
Total code size (KB) 120 289 514 508 483 578 979
First class size (KB) 5 1 11 2.5 4 7 1.5
Loading classes 0.2s 0.3s 0.5s 0.9s 1.0s 0.8s 1.3s
Parsing bytecodes 0.1s 0.1s 0.2s 0.5s 0.5s 0.4s 1.2s
Hashing bytecodes 0.1s 0.5s 0.7s 1.1s 1.4s 2.6s 2.9s
Additional latency 0.4s 0.9s 1.4s 2.5s 2.9s 3.8s 5.6s

Table 2. Breakdown of additional startup latency incurred by Deeds.

5 Discussion

History-based access-control for native binaries: History-based access-

control is not specific to Java or to mobile code. It can be used for any system
that allows interposition of code between untrusted programs and protected
resources. Several operating systems (Solaris, Digital Unix, IRIX, Mach and
Linux) allow users to interpose user-level code between an executing program
and OS resources by intercepting system calls. This facility is usually used to
implement debuggers and system call tracers. It has also been used to implement
 427

a general-purpose code interposition mechanism [24], a secure environment for

helper applications used by browsers to display files with different formats [11]
and a user-level file system [1]. It is also well-suited for implementing a Deeds-like
history-based access-control mechanism to mediate access to OS resources from
native binaries.
Pre-classified program behaviors: The one-out-of-k policy described in sec-
tion 2 classifies program behaviors in an on-line manner. A program gets classi-
fied as a browser, an editor, or a shell depending on whether it has connected to a
remote site, has opened local files for modification, or has created a sub-process.
To be able to do this for a wide variety of program behaviors, the policy that
does the classification and subsequent management of privileges has to contain
code to handle all these behaviors. An alternative scheme would be to allow
program-providers to label their programs with pre-classified behavior patterns
and to allow the users to specify which behaviors they would like to permit.
The policies governing individual behaviors could be added/deleted as need be.
While this scheme would require agreement on the labeling scheme, it is no more
complex than the MIME-types-based scheme that is already in use for display-
ing/processing different data formats. This scheme is similar to program-ACLs
and related defenses proposed for trojan-horse attacks [25,41].
Joint-authorization: Commercial applications, such as contracts and purchase
orders, may require multiple authorizations since the organization may wish to
reduce the risk of malfesance by dispersing trust over several individuals. History-
based access-control policies can be used to implement joint-authorization [39]
or k-out-of-n-authorizations [3]. For example, a policy may require that three out
of five known individuals must make the same request within the last T units of
time for the request to be granted; else the request is denied. In this case, the
history consists of the requests that have been made in the last T units of time.

6 R e l a t e d work

The primary problem for access-control mechanisms for mobile code is to be

able to differentiate between different programs executing on behalf of the same
user and to provide them with different privileges based on their expected be-
havior and/or potential to cause damage. A similar problem occurs in the con-
text of trojan-horse programs and viruses. To deal with such programs, sev-
eral researchers have developed mechanisms to limit the privileges of individual
programs based on their expected behavior [10,25-28,35,41]. Karger [25] uses
information about file extensions and behavior of individual programs to deter-
mine the set of files that a program is allowed to access (eg. a compiler invoked
on x . c is only allowed to create x. {o.u.out}). Lai [28] replaces the inference
mechanism by an explicit list of files accessible by a program. Wichers et al [41]
associate program-ACLs with each file thereby limiting the set of programs that
can access each file. King [26] uses a regular-expression-based language to specify
the set of objects each operation can access. Ko et al[27] use a language based
 428

on predicate logic and regular expressions to specify the security-relevant be-

havior of privileged programs and propose to use this specification for intrusion
detection. All these approaches assume that the set of programs to be run are
fixed and their behaviors are known. The mobile code environment is different
as the set of programs that will execute is inherently unknown. History-based
access-control is able to classify programs in an on-line manner and to thereafter
execute them within an environment with appropriate privileges. For example,
the one-out-of-k policy dynamically classifies downloaded programs into one of
three classes: browsers, editors and shells.
Schneider proposes EM as the class of security enforcement mechanisms
which work by monitoring the execution of a target (untrusted application in our
terminology) and which terminates any execution that is about to violate the se-
curity policy being enforced [34]. Behavior-based confinement is a part of the EM
class. Schneider also proposes a formalism (security automata) for specifying se-
curity policies. The finite-state nature of security automata implies that request-
histories for policies specified using them can be summarized using a counter.
This is similar to our experience with Deeds where we found that request-histories
for policies of interest could be summarized using booleans/counters.
The use of secure hash functions to derive a content-based name for software
has been proposed by Hollingsworth et al [16]. They propose to use these names
for configuration and version management of large applications and application
suites (such Microsoft Office).
An important feature of Deeds is its capability to install and compose mul-
tiple user-specified policies. Several researchers have proposed languages to al-
low users to specify access-control policies and frameworks to compose these
pohcies [3,13,18,19]. Three of them [13,18,19], propose logic-based declarative
languages and use inference mechanisms of various sorts to compose policies.
Blaze et al [3] propose a language that contains both assertions and procedural
filters and use a mechanism similar of that used in Deeds to implement com-
position. Access-control policies for Deeds are entirely procedural. Furthermore,
they can be updated while the programs whose accesses are being controlled are
still executing.
Two research groups have recently proposed constraint languages for speci-
fying security policies temporal aspects. Simon&Zurko [37] propose a language
for specifying temporal constraints such as HasDone, NeverDid, NeverUsed and
SomeoneFromEachfor separation of duty in role-based environments. These pred-
icates correspond to summaries of event-histories in Deeds terminology.
Mehta&Sollins [29] have independently proposed a constraint language for
specifying simple history-based access-control policies for Java applets. This
work was done in parallel with ours [7]. The approach presented in their pa-
per has two major limitations. First, they use the domain name of the server
that provides the applet as its identifier. This assigns the same identifier to all
applets from the same host. In addition, it is vulnerable to DNS spoofing at-
tacks. They suggest that this problem can be fixed by using the identity of the
author/supplier of an applet as its name. This assigns the same identifier to all
 429

applets from a single author/supplier and results in a single merged history. It

is not clear how such a merged history would be used as the predicates and
variables in their language are applet-specific. Even if each supplier provides
only one applet, this is a viable solution only if all the classes referenced by the
applet are provided in a single j ar file or are required to be signed by the same
principal. Otherwise, it is possible for a maHcious server that is able to spoof IP
addresses to intercept intermediate requests for dynamically linked classes and
provide malicious substitutes. The second limitation of their approach is that it
provides a small and fixed number of events. This limits the variety and power
of the policies that can be developed.
The event model used in Deeds is similar to that used in the SPIN extensible
operating system [2]. An interesting feature of SPIN is the use of dynamic com-
pilation to improve the performance of event dispatching [5]. If the performance
of event dispatching becomes a problem for Deeds (eg. if individual events have
a large number of handlers) we can use a similar technique.

7 C u r r e n t status and future directions

Deeds is currently operational and can be used for stand-alone Java programs.
We are in the process of identifying a variety of useful patterns of behaviors
and evaluating the performance and usability of Deeds in the context of these
behaviors.
In the near term, we plan to develop a history-based mechanism for me-
diating access to OS resources from native binaries. We also plan to explore
the possibility of using program labels to indicate pre-classified behaviors and
automatic loading/unloading of access-control policies to support this.
In the longer term, we plan to explore just-in-time binary rewriting to insert
event generation and dispatching code into downloaded programs. This would
allow users to create new kinds of events as and when they desire. Currently,
new kinds of events are created only by system libraries.

Acknowledgments

We would like to thank anonymous referees for their insightful comments which
helped improve the presentation of this paper.

References
1. A. Alexandrov, M. Ibel, K. Schauser, and C. Scheiman. Extending the operating
system at the user level: the Ufo global file system. In Proceedings of the 1997
USENIX Annual Technical Conference, 1997.
2. B. Bershad, S. Savage, P. Pardyak, et al. Extensibility, safety and performance
in the spin operating system. In Proc of the 15th ACM Symposium on Operating
System Principles, pages 267-84, 1995.
 430

3. M. Blaze, J. Feigenbaum, and J. Lacy. Decentralized trust management. In Proc

of the 17th Symposium on Security and Privacy, pages 164-73, 1996.
4. D. Brewer and M. Nash. The Chinese Wall Security Policy. In Proceedings of the
1989 IEEE Symposium, on Security and Privacy, 1989.
5. C. Chambers, S. Eggers, J. Auslander, M. Philipose, M. Mock, and P. Pardyak. Au-
tomatic dynamic compilation support for event dispatching in extensible systems.
In Workshop on Compiler Support for Systems Software, 1996.
6. B. Christiansen, P. Cappello, M. lonescu, M. Neary, K. Schauser, and D. Wu.
Javelin: Internet-based parallel computing using Java. In Proceedings of the 1997
ACM Workshop on Java for Science and Engineering Computation, 1997.
7. G. Edjlali, A. Achaiya, and V. Chaudhary. History-based access control for mobile
code. Technical report. University of California, Santa Barbara, 1997.
8. R. Englander. Developing Java Beans. O'Reilly & Associates, 1997.
9. J. Pritzinger and M. Mueller. Java security. Technical report. Sun Microsystems,
Inc, 1996.
10. T. Gamble. Implementing execution controls in Unix. In Proceedings of the 7th
System Administration Conference, pages 237-42, 1993.
11. I. Goldberg, D. Wagner, R. Thomeis, and E. Brewer. A secure environment for
untrusted helper applications: confining the wily hacker. In Proceedings of the
1996 USENIX Security Symposium, 1996.
12. L. Gong. New security architectural directions for Java. In Proceedings of IEEE
COMPCON'97, 1997.
13. C. Gunter and T. Jim. Design of an application-level security infrastructure. In
DIM ACS Workshop on Design and Formal Verification of Security Protocols, 1997.
14. The HashJava code obfuscator. Available from 4thPass Software,810 32nd Avenue
South, Seattle, WA 98144^
15. J. HoUingsworth, B. Miller, and J. Cargille. Dynamic program instrumentation for
scalable performance tools. In SHPCC, 1994.
16. J. HoUingsworth and E. Miller. Using content-derived names for caching and
software distribution. In Proceedings of the 1997 ACm Symposium on Software
Reusability, 1997.
17. C. Horstmann and G. Cornell. Core Java 1.1, volume I - Fundamentals. Sun
Microsystems Press, third edition, 1997.
18. T. Jaeger, A. Prakash, and A. Rubin. Building systems that flexibly control down-
loaded executable context. In Proc of the 6th Usenix Security Symposium, 1996.
19. S. Jajodia, P. Samarati, V. Subrahmanian, and E. Bertino. A unified framework
for enforcing multiple access control policies. In Proc. ACM SIGMOD Int'l. Conf.
on Management of Data, pages 474-85, 1997.
20. The JavaCC parser generator. Available from Sun Microsystems Inc. 901 San
Antonio Road, Palo Alto, CA 94303 USA^
21. The JaWavedit Audio File Editor. Available from Florian Bomers' web site^.
22. The Jeevan object-oriented database. Available from W3apps Inc., Ft. Lauderdale,
Florida''.
23. The JLex lexical analyzer generator. Available from the Department of Computer
Science, Princeton University*.
4
http://www.sbktech.org/hashjava.html
^ http://www.suntest.com/JavaCC
http://rummelplatz. uni-mannheim. de/ boemers/Ja Wavedit
http://www. w3apps.com
http://www. cs.princeton. edu/ appel/modern/java/JLex
 431

24. M. Jones. Interposition agents: Transparently interposing user code at the sys-
tem interface. In Proceedings of the 14th ACM Symposium on Operating System
Principles, 1993.
25. P. Karger. Limiting the damage potential of the discretionary trojan horse. In
Proceedings of the 1987 IEEE Syposium on Research in Security and Privacy, 1987.
26. M. King. Identifying and controlling undesirable program behaviors. In Proceedings
of the 14th National Computer Security Conference, 1992.
27. C. Ko, G. Fink, and K. Levitt. Automated detection of vulnerabilities in privileged
programs by execution monitoring. In Proceedings. 10th Annual Computer Security
Applications Conference, pages 134-44, 1994.
28. N. Lai and T. Gray. Strengthening discretionary access controls to inhibit tro-
jan horses and computer viruses. In Proceedings of the 1988 USENIX Summer
Symposium, 1988.
29. N. Mehta and K. SoUins. Extending and expanding the security features of Java.
In Proceedings of the 1998 USENIX Security Symposium, 1998.
30. Microsoft Corporation. Proposal for Authenticating Code Via the Internet, Apr
1996. http://www.microsoft.com/intdev/security/authcode.
31. R. Rivest. The MD5 message-digest algorithm. RFC 1321, Network Working
Group, 1992.
32. J. Saltzer and M. Schroeder. The protection of information in computer systems.
Proceedings of the IEEE, 63(9):1278-1308, Sep 1975.
33. R. Scheifler and J. Gettys. X Window System : The Complete Reference to Xlib,
X Protocol, Icccm, Xlfd. Butterworth-Heinemann, 1992.
34. F. Schneider. Enforceable security policies. Technical report, Dept of Computer
Science, Cornell University, 1998.
35. C. Serban and B. McMillin. Run-time security evaluation (RTSE) for distributed
applications. In Proc. of the 1996 IEEE Symposium on Security and Privacy, pages
222-32, 1996.
36. Secure hash standard. Federal Information Processing Standards Publication,
FIPS, PUB 180-1, April 1995.
37. R. Simon and M. Zurko. Separation of duty in role-based environments. In Pro-
ceedings of the IEEE Computer Security Foundations Workshop '97, 1997.
38. The Spaniel News Server. Available from Spaniel Software^.
39. V. Varadharajan and P. Allen. Joint actions based authorization schemes. Oper-
ating Systems Review, 30(3):32-45, 1996.
40. D. Wallach, D. Balfanz, D. Dean, and E. Felten. Extensible security architecture
for Java. In SOSP 16, 1997.
41. D. Wichers, D. Cook, R. Olsson, J. Crossley, P. Kerchen, K. Levitt, and R. Lo.
PACL's: an access control list approach to anti-viral security. In USENIX Work-
shop Proceedings. UNIX SECURITY II, pages 71-82, 1990.
42. The WingDis Editor. Available from WingSoft Corporation, P.O.Box 7554, Fre-
mont, CA 94537^°.

9
http://www.searchspaniel.com/newsserver.html
10
http://www.wingsoft.com/javaeditor.shtml
 Security in Active Networks

D. Scott Alexander, William A. Arbaugh, Angelos D. Keromytis, and

J o n a t h a n M. Smith

Abstract. The desire for flexible networking services has given rise to
the concept of "active networks." Active networks provide a general
framework for designing and implementing network-embedded services,
typically by means of a programmable network infrastructure. A pro-
grammable network infrastructure creates significant new challenges for
securing the network infrastructure.
This paper begins with an overview of active networking. It then moves
to security issues, beginning with a threat model for active networking,
moving through an enumeration of the challenges for system designers,
and ending with a survey of approaches for meeting those challenges.
The Secure Active Networking Environment (SANE) realizes many of
these approaches; an implementation exists and provides acceptable per-
formance for even the most aggressive active networking proposals such
as active packets (sometimes called "capsules").
We close the paper with a discussion of open problems and an attempt
to prioritize them.

1 W h a t is Active Networking?

In networking architectures a design choice can be made between:

1. Restricting the actions of the network infrastructure t o t r a n s p o r t , and

2. easing those restrictions to permit on-the-fly customization of the network
infrastructure.

T h e d a t a - t r a n s p o r t model, which has been successfully applied in the I P Internet

and other networks, is called passive networking since the infrastructure {e.g., IP
routers) is mostly indifferent t o the packets passing through, and their actions
(forwarding and routing) cannot be directly influenced by users. This is not t o
say t h a t the switches do not perform complex computations as a result of re-
ceiving or forwarding a packet. Rather, the nature of these computations cannot
dynamically change beyond the fairly basic configuration options provided by
the manufacturer of the switch.
In contrast, active networking allows network-embedded functionality other
t h a n t r a n s p o r t . For current systems, this functionality ranges from W W W proxy
caches, multicasting [17] and R S V P [12] to firewalls. Since each of these indepen-
dently designed and supported functions could be carried out as an application
of a more general infrastructure, the architecture of such active infrastructures
is now being investigated aggressively.
 434

The basic principle employed is the use of programmability, as this allows

many applications to be created, including those not foreseen by the designers
of the switch. There are a number of forms this programmability can take, in-
cluding treating each packet as a program (active packets or "capsules") and
programming or reprogramming network elements on-the-fly with select pack-
ets. Note that the latter approach subsumes the former, as a program may be
loaded that treats all subsequent packets as programs.

1.1 W h y is Active Network Security Interesting?

Prom a security perspective, a large scale infrastructure with user access to

programming capabilities, even if restricted, creates a wide variety of difficult
challenges. Most directly, since the basis of security is controlled access to re-
sources, the increased complexity of the managed resources makes securing them
much more difficult. Since "security" is best thought of as a mapping between
a pohcy and a set of predicates maintained through actions, the pohcy must be
more complex than, in as much as they exist, equivalent policies of present-day
networks, resulting in an explosion in the set of predicates.
For example, the ability to load a new queuing discipline may be attractive
from a resource control perspective, but if the queuing discipline can replace that
of an existing user, the replacement policy must be specified, and its implemen-
tation carefully controlled through one or more policy enforcement mechanisms.
Additionally, such a scenario forces the definition of principals and objects
with which policies are associated. When compared with the policy at a basic
IP router (no principals, datagram dehvery guarantees, PIFO queuing, etc.) it
can be seen why securing active networks is difficult.

1.2 Virtual and Real Resources

As the role of active networking elements is to store, compute and forward, the
managed resources are those required to store packets, operate on them, and
forward them to other elements. The resources provided to various principals
at any instant cannot exceed the real resources {e.g., output port bandwidth)
available at that instant. This emphasis on real resources and time implies that
a conventional < object, principal, access> 3-tuple for an access control list (ACL)
is inadequate.
To provide controlled access to real resources, with real time constraints, a
fourth element to represent duration (either absolute or periodic) must be added,
giving <object, principal, access, QoS guarantees>. This remains an ACL, but is
not "virtualized" by leaving time unspecified and making "eventual" access ac-
ceptable. We should point out that this new element in the ACL can be encoded
as part of the access field. Similarly, we need not use an actual ACL, but we may
use mechanisms that can be expressed in terms of ACLS and are better-suited
for distributed systems.
 435

2 Terminology
The term trust is used heavily in computer security. Unfortunately, the term
has several definitions depending on who uses it and how the term is used. In
fact, the U.S. Department of Defense's Orange Book [20], which defined several
levels of security a computer host could provide, defines trust ambiguously. The
definition of trust used herein is a slight modification of that by Neumann [46].
An object is defined as trusted when the object operates as expected according to
design and policy. A stronger trust statement is when an object is trustworthy. A
trustworthy object is one that has been shown in some convincing manner, e.g.,
a formal code-review or formal mathematical analysis, to operate as expected. A
security-critical object is one which the security — defined by a policy — of the
system depends on the proper operation of the object. A security-critical object
can be considered trusted, which is usually the case in most secure systems, but
unfortunately this leads to an unnecessary profusion of such objects.
We note the distinction between trust and integrity: Trust is determined
through the verification of components and the dependencies among them. In-
tegrity demonstrates that components have not been modified. Thus integrity
checking in a trustworthy system is about preserving an established trust or
trust relationship.

2.1 Threat Model

An active network infrastructure is very different from the current Internet [3].
In the latter, the only resources consumed by a packet at a router are:
1. the memory needed to temporarily store it, and
2. the CPU cycles necessary to find the correct route.
Even if IP [53] option processing is needed, the CPU overhead is still quite small
compared to the cost of executing an active packet. In such an environment, strict
resource control in the intermediate routers was considered non-critical. Thus,
security policies [8] are enforced end-to-end. While this approach has worked
well in the past, there are several problems. First, denial-of -service attacks are
relatively easy to mount, due to this simple resource model. Attacks to the
infrastructure itself are possible, and result in major network connectivity loss.
Finally, it is very difficult to provide enforceable quality-of-service guarantees.
[12]
Active Networks, being more flexible, considerably expand the threat possi-
bilities, because of the increased numbers of potential points of vulnerability. For
example, when a packet containing code to execute arrives, the system typically
must:

— Identify the sending network element.

— Identify the sending user.
— Grant access to appropriate resources based on these identifications.
— Allow execution based on the authorizations and security policy.
 436

In networking terminology, the first three steps comprise a form of admission

control, while the final step is a form of policing. Security violations occur when
a policy is violated, e.g., reading a private packet, or exceeding some specified
resource usage. In the present-day Internet, intermediate network elements {e.g.,
routers) very rarely have to perform any of these checks. This is a result of the
best-effort resource allocation policies inherent in IP networking.

Denial-of-Service Attacks. Cryptographic mechanisms have proven remark-

ably successful for functions such as identification and authentication. These
functions typically (although not necessarily) are used in protocols with a vir-
tual time model, which is concerned with sequencing of events rather than more
constrained sequencing of events with time limits (the real time model). The
cases where time limits are observed are almost always for reasons of robust-
ness, e.g., to force eventual termination. Since such timeouts are intended for
extreme circumstances, they are long enough so that they can cope with any
reasonable delay.
In an environment where a considerable fraction (and perhaps eventually a
majority) of the traffic will be continuous media traffic, security must include re-
source management and protection with an eye to preserving timing properties.
In particular, a pernicious form of "attack" is the so-called "denial-of-service" at-
tack. The basic principle applied in such an attack is that while wresting control
of the service is desirable, the goal can be achieved if the opponent cannot use
the service. This principle has been used in military communications strategies,
e.g., the use of radio "jamming" to frustrate an opponent's communications,
and most recently in denying service to Internet Service Provider servers using
a TCP 51W flood attack [50, 16]. Another very effective (even crippling) attack
on a computer system can occur due to scheduling algorithms which implicitly
embed design assumptions.
To look at an example in some detail, consider the so-called "recursive shell"
shown in Figure 1.
The shell script invokes itself. This is in fact a natural programming style,
except that the process of invoking a shell script consists mainly of executing
two heavyweight system calls, f o r k O and exec(), which, respectively, create
a new copy of the current process and replace the current process with a new
process created from an executable file. Since the program spends the majority
of its time executing system calls, which in UNIX cause the operating system
to execute on behalf of the user (at high priority) the system's resources are
typically consumed by this program (including CPU time and table space used
for holding process control blocks).
With an active network element, it is easy to imagine situations where user
programs (or errant system programs) run amok, and make the network elements
useless for basic tasks. The solution, we believe, is to constrain real resources
associated with active network programs. For example, if we limited the principal
{e.g., a "user") invoking the recursive shell script to 10% of the CPU time, or
10% of the system memory, the process would either limit its effects on the CPU
 437

to a 10% degradation, or fail to operate (since it could not invoke a new process)
when it hit the table space limitation. Fortunately, a number of new operating
systems [40, 35] have appeared which provide the services necessary to contain
one or more executing threads within a single scheduling domain.

#!/bin/sh
$0 #invoke ourselves

Fig. 1. A recursive shell script for UNIX

2.2 Challenges for the System Designer

Independent of the specific network architecture, the designer of a network has

a set of tradeoffs they must make which define a "design space." We consider
five here:

1. Flexibility. Flexibility is a measure of the system to perform a variety of

tasks.
2. Usability. Usability is a measure of the ease with which the system can be
used for its intended task(s).
3. Performance. The system will have some quantitative measures by which it
is evaluated, such as throughput, delay, delay variation.
4. Cost. A networking system will have quantifiable economic costs, such as
costs for construction, operation, maintenance and continuing improvements.
5. Security. Since network systems are shared resources the designer must pro-
vide mechanisms to protect users from each other according to a policy.

It is our belief that, as in this list, security is often left until last in the
design process, which results in not enough attention and emphasis being given
to security. If security is designed in, it can simply be made part of the design
space in which we search for attractive cost/performance tradeoffs. For example,
if acceptable flexibility requires downloadable software, and acceptable security
means that only trusted downloadable software will be loaded, our cost and
performance optimizations will reflect ideas such as minimizing dynamic checks
with static pre-checks or other means. If security is not an issue, there is no
point in doing this.
The designer's major challenge is finding a point (or set of points) in the
design space which is acceptable to a large enough market segment to influence
the community of users. Sometimes this is not possible; the commercial empha-
sis on forwarding performance is so overwhelming that concessions to security
slowing the transport plane are simply unacceptable. Fortunately, organizations
have become sufficiently dependent on information networks that security does
sell.
 438

In the context of active networks, the major focus of security is the set of
activities which provide flexibihty; that is, the faciUty to inject new code "on-
the-fly" into network elements. To build a secure infrastructure, first, the in-
frastructure itself (the "checker") must be unaltered. Second, the infrastructure
must provide assurance that loaded modules (the dynamic checking) will not
violate the security properties. In general, this is very hard. Some means cur-
rently under investigation include domain-specific languages which are easy to
check {e.g., PLAN), proof-carrying code [45, 44], restricted interfaces (ALIEN),
and distributed responsibility (SANE). Currently, the most attractive point in
the design space appears to be a restricted domain-specific language coupled to
an extension system with heavyweight checks. In this way, the frequent (per-
packet) dynamic checks are inexpensive, while focusing expensive scrutiny on
the extension process. This idea is manifest in the Switch Ware active network
architecture [2].

2.3 Possible Approaches

Security of Active Networks is a broad evolving area. We will mention only
some of the most directly relevant related work. In addition to the related works
sections of the papers listed, we suggest Moore [41] as a source of additional
information in this area.
Software fault isolation as a safety mechanism for mutually-suspicious mod-
ules running in the same address space was introduced in [62]. This technique
involves inserting run-time checks in the application code. While it has been
successfully demonstrated for RISC architectures, application of the same tech-
niques to CISC architectures remains problematic.
Typed assembly language [42] propagates type safety information to the as-
sembly language level, so assembly code can be verified. However, there are sev-
eral security properties {e.g., resource usage, which is a dynamic measure) that
do not easily map into the type-checking model because of the latter's static
nature.
Proof-carrying code [44] permits arbitrary code to be executed as long as a
valid proof of safety accompanies it. While this is a very promising technique, it
is not clear that all desirable security properties and policies are expressible and
provable in the logic used to publish the policy and encode the proof. Used in
conjunction with other mechanisms, we believe that it will prove a very useful
security tool.
PLAN [27, 28] is a part of the Switch Ware [2, 57] project at the University of
Pennsylvania. The PLAN project is investigating the tradeoffs brought about by
using a different language for active packets than is used for active extensions.
They have designed a new language called PLAN (which is loosely based on ML
[39]). PLAN is designed so that pure PLAN programs will not be able to violate
the security policy. This policy is intended to be sufficiently restrictive that node
administrators will be willing to allow PLAN programs to run without requiring
authentication. Because this limits the operations that can be performed, PLAN
programs can call services which can either be active extensions or facilities built
 439

into the system. These services may require authentication and authorization
before allowing access to the resources they protect.
The Safetynet Project [63] at the University of Sussex has also designed a new
language for active networking. They have explicitly enumerated what they feel
are the important requirements for an active networking language and then set
about designing a language to meet those requirements. In particular, they differ
from PLAN in that they hope to use the type system to allow safe accumulation
of state. They appear to be trying to avoid having any service layer at all.
Java [23] and ML [39, 34] (and the MMM [37] project) provide security
through language mechanisms. More recent versions of Java provide protection
domains [22]. Protection domains were first introduced in Multics [55, 56, 38, 54].
These solutions are not applicable to programs written in other languages (as
may be the case with a heterogeneous active network with multiple execution
environments), and are better suited for the applet model of execution than
active networks. The need for a separate bytecode verifier is also considered
by some a disadvantage, as it forces expensive (in the case of Java, at lea^t)
language-compliance checks prior to execution. In this area, there is some re-
search in enhancing the understanding of the tradeoffs between compilation
time/complexity, and bytecode size, verification time, and complexity.
It should be noted that language mechanisms can (and sometimes do) serve as
the basis of security of an active network node. Other language-based protection
schemes can be found in [9, 13, 26, 36, 33, 24].

3 SANE Architecture

Previous attempts at system security have not taken a holistic approach. The
approaches typically focused on a major component of the system. For instance,
operating system research has usually ignored the bootstrap process of the host.
As a result, a trustworthy operating system is started by an untrustworthy boot-
strap! This creates serious security problems since most Operating Systems re-
quire some lower level services, e.g., firmware, for trustworthy initiahzation and
operation. A major design goal of SANE [3] was to reduce the number and size
of components that are assumed as trustworthy. A second major design goal of
SANE was to provide a secure and reliable mechanism for establishing a secu-
rity context for active networking. An application or node could then use that
context in any manner it desired.
No practical system can avoid assumptions, however, and SANE is no dif-
ferent. Two assumptions are made by SANE. The first assumption is that the
physical security of the host is maintained through strict enforcement of a phys-
ical security policy. The second assumption SANE makes is the existence of a
Public Key Infrastructure (PKI). While a PKI is required, no assumptions are
made as to the type of PKI, e.g., hierarchical or "web of trust."[15, 31, 66, 10,11]
The overall architecture of SANE for a three-node network is shown in Fig-
ure 2.
 440

The initialization of each node begins with the bootstrap. Following the
sucessful completion of the bootstrap, the operating system is started which
loads a general purpose evaluator, e.g., a Caml [34] or Java [23] runtime. The
evaluator then starts an "Active Loader" which restricts the environment pro-
vided by the evaluator. Finally, the loader loads an "Active Network Evaluator"
(ANE) which accepts and evaluates active packets, e.g., PLAN [27], Switchlet,
or ANTS [64]. The ANE then loads the SANE module to estabhsh a security
context with each network neighbor. Following the establishment of the security
context, the node is ready for secure operation within the active network.
It should be noted that the services offered by SANE can be used by most
active networking schemes. In our current system, SANE is used in conjunction
with the ALIEN architecture [1]. ALIEN is built on top of the Caml runtime,
and provides a network bytecode loader, a set of libraries, and other facilities
necessary for active networking.
The following sections describe the three components of SANE. These include
the AEGIS [5, 6] bootstrap system, the ALIEN [1] architecture, and SANE [2, 3]
itself.

3.1 AEGIS Bootstrap

AEGIS [5] modifies the standard IBM PC process so that all executable code,
except for a very small section of trustworthy code, is verified prior to execution
by using a digital signature. This is accomplished through modifications and ad-
ditions to the BIOS (Basic Input/Output System). In essence, the trustworthy
software serves as the root of an authentication chain that extends to the eval-
uator and potentially beyond, to "active" packets. In the AEGIS boot process,
either the Active Network element is started, or a recovery process is entered to
repair any integrity failure detected. Once the repair is completed, the system
is restarted to ensure that the system boots. This entire process occurs with-
out user intervention. AEGIS can also be used to maintain the hardware and
software configuration of a machine.
It should be noted that AEGIS does not verify the correctness of a software
component. Such a component could contain an exploitable fiaw. The goal of
AEGIS is to prevent tampering of components that are considered trustworthy
by the system administrator. AEGIS verifies the integrity of already trusted
components. The nature of this trust is outside the scope of this paper.
Other work on the subject of secure bootstrapping includes [59, 65,14, 32, 25].
A more extensive review of AEGIS and its differences with the above systems
can be found in [5, 6].

AEGIS Layered Boot and Recovery Process. AEGIS divides the boot
process into several levels to simplify and organize the BIOS modifications, as
shown in Figure 3. Each increasing level adds functionality to the system, pro-
viding correspondingly higher levels of abstraction. The lowest level is Level 0.
Level 0 contains the small section of trustworthy software, digital signatures.
 441

Opcniling System
Operating System
Y
Libraries SAX SAX Libraries
_A_ . /^ -
f Switchlet j
r Switchlet \ *

( Active Loader J c Active Loader )

Caml / Java
Bootstrap

Active Packets Security Association Exchange (SAX)

F i g . 2 . S A N E Network A r c h i t e c t u r e

public key certificates, and recovery code. The integrity of this level is assumed
as valid. We do, however, perform an initial checksum test to identify PROM
failures. The first level contains the remainder of the usual BIOS code and the
CMOS. The second level contains all of the expansion cards and their associated
ROMs, if any. The third level contains the operating system boot sector. These
are resident on the bootable device and are responsible for loading the operating
system kernel. The fourth level contains the operating system, and the fifth and
final level contains the ALIEN architecture and other active nodes..
The transition between levels in a traditional boot process is accomplished
with a jump or a call instruction without any attempt at verifying the integrity
of the next level. AEGIS, on the other hand, uses public key cryptography and
cryptographic hashes to protect the transition from each lower level to the next
higher one, and its recovery process through a trusted repository ensures the
integrity of the next level in the event of failures [6].
 442

The trusted repository can either be an expansion ROM board that contains
verified copies of the required software, or it can be another Active node. If the
repository is a ROM board, then simple memory copies can repair or shadow
failures. In the case of a network host, the detection of an integrity failure causes
the system to boot into a recovery kernel contained on the network card ROM.
The recovery kernel contacts a "trusted" host through the secure protocol de-
scribed in [6, 7] to recover a signed copy of the failed component. The failed
component is then shadowed or repaired, and the system is restarted (warm
boot).

W yXctive
rsJetwork
£^lement

•4 ,

f
^

Af ^ ^ 1 "

— — — —J _.^^^ p^^.f^

Fig. 3. AEGIS boot control flow

3.2 The ALIEN Architecture

The basis of the ALIEN approach is that we take a general model of computa-
tion and restrict it. Caml provides the general model and ALIEN provides the
restrictions.
More precisely, Caml provides a model of computation equivalent to that of
a Turing machine. By itself, this computation model is secure since it involves
no shared resources. In practice, since we are running on a real machine, we have
denial-of-service attacks that arise because our CPU and memory resources are
finite. Additionally, the actual Caml environment also includes a runtime sys-
tem that, among other features, provides access to operating system primitives.
 443

which, in turn, provide access to shared resources. Further, under this runtime,
memory is a shared resource. The role of ALIEN is to control the access to these
shared resources and thereby ensure that a loaded program (called "switchlet")
does not exceed its resource limits (ALIEN is not responsible for determining
those hmits).
ALIEN itself is built of three major components. The Loader provides the
interface to the Objective Caml runtime system. The Core Switchlet builds on
the Loader both by providing the security-related restrictions required and by
providing more generally useful interfaces to low-level functions. Finally, the
libraries are sets of utility routines. Each of these pieces will be briefly covered
in turn in the following paragraphs.

The Loader. The Loader provides the core of ALIEN's functionality. It provides
the interface to the operating system (through the language runtime) plus some
essential functions to allow system startup and loading of switchlets, as shown in
Table 1. Thus, it defines the "view of the world" for the rest of ALIEN. Moreover,
since security involves interaction with either the external system or with other
switchlets, the Loader provides the basis of security.
It should be noted that Loader provides mechanisms rather than pohcy;
pohcies in the Core Switchlet can be changed by changing pieces of the Core
Switchlet.

startup routines initialize system

switchlet loading dynamically load switchlets consistent with ALIEN security
system console console read loop
Table 1. Loader functionality

The Core Switchlet. Above the Loader is the Core Switchlet. It is responsible
for providing the interface that switchlets see. It relies upon the Loader for
access to operating system resources, and then layers additional mechanisms
to add security and, often, utility. In providing an interface to switchlets, it
determines the security policies of the system. By including or excluding any
function, it can determine what switchlets can or cannot do. Since it is loadable,
the administrator can change or upgrade its pieces as necessary. This also allows
for changes in the security policy.
The policies of the Core Switchlet are enforced through a combination of
module thinning and type safety. Type safety ensures that a switchlet can only
access data or call functions that it can name. This allows implementations of
ALIEN that run in a single address space, thus avoiding the overheads normally
associated with crossing hardware-enforced boundaries. [47].
Module thinning allows the Core Switchlet to present a limited interface to
switchlets. Combining this with type safety, switchlets can be prevented from
 444

calling functions or accessing data even though they share an address space.
It is even possible to differentiate switchlets so as to provide a rich interface
to a trusted switchlet or to provide a very limited interface to an anonymous
switchlet. Similar approaches have been taken in [33, 9, 61].
In many ways, the interface that the Core Switchlet presents to switchlets and
libraries is like the system call interface that a kernel presents to applications.
Through design of the interface the system can control access to underlying
resources. With a well-designed interface, the caller can combine the functions
provided to get useful work done. Table 2 shows the functionality provided by
the Core Switchlet.

language primitives policy for access to the basic functions of the language
operating system access policy for access to the operating system calls
network access policy and mechanism for access to the network
thread access policy for access to threads primitives
loading support policy and mechanism to support loading of switchlets
message logging policy and mechanism for adding messages to the log file

Table 2. Core Switchlet functionality

Because we are implementing a network node, access to the network is par-

ticularly important. Generally this consists of allowing switchlets to discover
information about the interfaces on the machines and the attached networks,
receive packets, and send packets. One element of this task which is particularly
important is the demultiplexing of incoming packets. The Core Switchlet must
be able to determine whether zero, one, or more than one switchlet is interested
in an arriving packet. If more than one switchlet is interested in the packet,
policy should dictate which switchlet or switchlets receive a copy of the packet.
Security is an important element of the decision as a switchlet should be able
to be certain that it will get all packets that it should receive under the policy,
and should not be able to get any packets that it should not receive under the
policy. Without such security, denial-of-service attacks and information stealing
are quite easy.

The Librciry. The library is a set of functions which provide useful routines
that do not require privilege to run. The proper set of functions for the library is
a continuing area of research. Some of the things that are in the library for the
experiments we have performed include utility functions and implementations of
IP and UDP [52].

Locating Functionality. When expanding our implementation, it is not always

obvious in which layer the new functionality belongs. In this section, we present
the principles we use to make this determination. Our first principle is that if the
 445

functionality can be implemented in a library, it should be. Said another way,

if the functions exposed by the Core Switchlet or available from other libraries
provide the infrastructure needed to implement the new functionality, a library
is warranted.
If the new functionality relies on some element of the runtime not made avail-
able to unprivileged code, then either the Loader or the Core Switchlet must be
expanded. Because these elements define the common, expected interface avail-
able at the switch, we attempt to keep them small to minimize the required
resources. Therefore, our second principle is that we prefer to break off the
smallest reasonable portion of the new functionality (consistent with security)
that can be implemented in the privileged parts of the system. The remainder
becomes a library. In our experience this also aids generality, as the privileged
portion is often useful to other libraries developed later. For example, to imple-
ment IP, we built a small module inside the Core Switchlet which reads Ethernet
frames from the operating system. It also demultiplexes the frames based on the
Ethernet type field to increase generality. The remainder of IP, which processes
headers, could then be made a non-privileged library.
Our third principle is that if this privileged functionality sets policy, it needs
to go into the Core Switchlet. As discussed above, policy must be set in the
Core Switchlet so that the loading mechanism can be used as needed to change
policy. Our final principal is any functionality that provides pure mechanism is
placed in the Core Switchlet unless it is needed before the Core Switchlet can
be running.

3.3 S A N E Services
SANE builds on AEGIS and ALIEN in order to provide security services for an
active network. We believe that these services are required for the deployment of
a robust active infrastructure. This is not to say that they contain all the security
mechanisms one would ever want. Rather, they are basic building blocks needed
for possibly more advanced mechanisms. These services include:

- Cryptographic primitives, provided as ALIEN hbraries. A number of sym-

metric [43] and public-key [48] cryptosystems and hash [49] functions are
made available for use by programmers and other system components. These
are building blocks for cryptographic protocols and other mechanisms that
provide higher-order security primitives.
- Packet authentication. This can be achieved through signing the packet with
a public-key algorithm, or using some secret key method, hke a MAC. Au-
thentication can be used in a number of different contexts:
• When dynamically loading switchlets over the network, to ensure code
integrity and proper access authorization/resource allocation.
• Similarly, when transmitting data over the network, to ensure data in-
tegrity, and packet-flow isolation. This latter may form the basis for
economic traffic-management schemes, if the authentication mechanisms
prove lightweight enough.
 446

Public-key authentication allows for zero-roundtrip authentication, since no

negotiation is required, but is relatively heavy-weight computationally and is
subject to some replay attacks in the absence of node-persistent state or syn-
chronized clocks [58, 21] in the network switches. This form of authentication
is well-suited to mobile-agent types of applications, such as some active net-
work management schemes [51], or where data generated on a switch needs
to be securely combined with the code.
Secret-key based authentication is faster, and thus better suited for bulk
transport. For scalability reasons however, it needs to be automated, and
thus needs some public-key infrastructure to base the authentication on.
The cost of the automated key establishment can be easily amortized after
transmitting even a small amount of data.
Packet confidentiality (encryption). The same issues with regards to public
vs. secret key authentication are present here. The uses of this service are
analogous to the packet authentication service.
A key establishment protocol (KEP), which allows two principals in the
network to establish secret keys and exchange certificates. The protocol is
also used in bootstrap failure-recovery in AEGIS [6] and is based on the
Station-to-Station [19] protocol, using Diffie-Hellman [18] key exchange and
DSA [48] (or other) public-key signatures. This protocol is used in three
different roles:
1. Secure boostrap component recovery in AEGIS, as we discussed in Sec-
tion 3.1.
2. Secure neighbor discovery once the node boots. Similar to verifying the
network cards and software components in AEGIS, at this stage the node
verifies the immediate network topology and establishes trust between
"adjacent" switches. The shared secret keys established in this manner
can be used for hop-by-hop secure transmission of data or code, which is
a requirement for some mobile-agent types of applications. Other critical
infrastructure information can be secured by this mechanism as well, e.g.,
routing updates.
3. Session-key establishment, principal authentication and authorization.
The shared secret keys can be used to secure the communications of any
two principals in the network. Furthermore, principals can authenticate
each other and exchange authorization credentials. For example, a user
can verify the identity of a switch he needs to load a program on, while
the switch can determine whether the user has permission to do so,
and what other restrictions apply. We make use of KeyNote [10, 11]
credentials to provide this functionality. These credentials specify the
resource usage and access control policies that ALIEN enforces.
Administrative domains allow a set of network elements under the same
administrative control to restrict security requirements when communicating
with each other. In such a configuration, "border" elements act as present-
day firewalls and can require e.g., a number of iterations of KEP to estabhsh
all the problem credentials. They can then mark the active packets such that
further negotiation in order to determine credentials is not needed by the
 447

"interior" switches. In essense, these "active firewalls" act as introducers of

"outsiders" in a closed system. Administrative domains are built on top of
secure packet exchange, in conjunction with the key management protocol.
Naming services allow for unsupervised but collision-free^ (secure) identi-
fication of programs. The basis of this approach is to combine hashes of
the code, public keys (and signatures), and user-defined strings to gener-
ate "names" for pieces of code. Thus, a certain program can have different
names, each with different semantics and trust dependencies. Such a service
is necessary in an active network environment where different users' modules
can explicitly interact with, or even depend on, each other. For more details,
see [3].

4 Conclusions and F u t u r e Work

This paper has presented an overview of the programmable network architectures

called "Active Networks" and illustrated the security issues facing the architects
of such systems. While the goal of Active Networks is to increase the set of
design options available in distributed systems, the flexibility of a programmable
infrastructure introduces considerable burdens for control and management. We
have outlined, and illustrated in some detail, strategies for coping with some of
these issues. Many issues currently remain unresolved, and are thus important
directions for future work.
First and foremost is the requirement that newly configured functionality not
damage the network as an aggregate, in addition to not damaging the network
element into which it is configured. The means for attacking such problems
remains unclear, but two of the more promising approaches are the use of formal
languages and methods for distributed concurrent systems (such as CSP [29, 30]),
and the use of economic methods to deal with aggregate costs which accrue in
a distributed fashion.
The second important direction is time-sensitive resource access, where Qual-
ity of Service (QoS) is part of the resource access model to which the security
architecture applies. Much of the fundamental work of computer security has
been based on a time-independent model of resource access, to which symbolic
logic and formal methods can be applied. However, as any user of information,
whether "secure" or "insecure" is well aware, if the information doesn't get there
in time, it is useless. If the service is not provided, it has failed. Thus we have to
push our security models to reflect time, so that service (or the attackers goal,
"denial-of-service") is as first class as identification or access control in network
security architectures.
Finally, we must worry about scale. Some of the solutions we have discussed
here work well in the small but must be automated to permit scaling to large
systems. For example, the SANE model of permitting resource access to trusted

^ To the extent that the cryptographic hash functions employed are resistant to colli-
sions.
 448

entities presumes trust establishment and trust specification. In a (multi-) mil-

lion node network, such trust management must be carried out automatically,
and must b e globally specified using human-comprehensible policies rather t h a n
node-node relationships.

5 Acknowledgements

Portions of this paper are u p d a t e d from [3] and [4]. This work was supported by
D A R P A under Contract #N66001-96-C-852, with additional support from the
Intel Corporation.

References
[1] D. S. Alexander. ALIEN: A Generalized Computing Model of Active Networks.
PhD thesis, University of Pennsylvania, September 1998.
[2] D. S. Alexander, W. A. Arbaugh, M. Hicks, P. Kakkar, A. D. Keromytis, J. T.
Moore, C. A. Gunter, S. M. Nettles, and J. M. Smith. The Switch Ware Active
Network Architecture. IEEE Network Magazine, special issue on Active and Pro-
grammable Networks, 12(3):29-36, 1998.
[3] D. S. Alexander, W. A. Arbaugh, A. D. Keromytis, and J. M. Smith. A Secure
Active Network Environment Architecture: Realization in SwitchWaxe. IEEE
Network Magazine, special issue on Active and Programmable Networks, 12(3):37-
45, 1998.
[4] D. Scott Alexander, Wilham A. Arbaugh, Angelos D. Keromytis, and Jonathan M.
Smith. Safety and Security of Programmable Network Infrastructures. IEEE
Communications Magazine, 36(10):84 - 92, 1998.
[5] W. A. Arbaugh, D. J. Farber, and J. M. Smith. A Secure and Reliable Bootstrap
Architecture. In Proceedings 1997 IEEE Symposium on Security and Privacy,
pages 65-71, May 1997.
[6] W. A. Arbaugh, A. D. Keromytis, D. J. Farber, and J. M. Smith. Automated
Recovery in a Secure Bootstrap Process. In Proceedings of Network and Distributed
System Security Symposium, pages 155-167. Internet Society, March 1998.
[7] W. A. Arbaugh, A. D. Keromytis, and J. M. Smith. DHCP-I-I-: Applying an effi-
cient implementation method for fail-stop cryptographic protocols. In Proceedings
of Global Internet (GlobeCom) '98, November 1998.
[8] R. Atkinson. Security Architecture for the Internet Protocol. RFC 1825, August
1995.
[9] B. Bershad, S. Savage, P. Pardyak, E. G. Sirer, M. Fiuczynski, D. Becker, S. Eg-
gers, and C. Chambers. Extensibility, safety and performance in the spin operating
system. In Proc. 15th SOSP, pages 267-284, December 1995.
[10] M. Blaze, J. Feigenbaum, J. loannidis, and A. Keromytis.
The KeyNote Trust-Management System. Work in Progress,
http://www.cis.upenn.edu/"angelos/keynote.html, June 1998.
[11] M. Blaze, J. Feigenbaum, J. loannidis, and A. Keromytis. The role of trust man-
agement in distributed systems security. In Secure Internet Programming [60].
[12] R. Braden, L. Zhsing, S. Berson, S. Herzog, and S. Jamin. Resource ReSerVation
Protocol (RSVP) - Version 1 Functional Specification. Internet RFC 2208, 1997.
 449

[13] J. S. Chase, H. M. Levy, M. J. Feeley, and E. D. Lazowska. Sharing and Protection

in a Single-Address-Space Operating System. In ACM TVansactions on Computer
systems, November 1994.
[14] Paul Christopher Clark. BITS: A Smartcard Protected Operating System. PhD
thesis, George Washington University, 1994.
[15] Consultation Committee. X.509: The Directory Authentication Framework. In-
ternational Telephone and Telegraph, International Telecommunications Union,
Geneva, 1989.
[16] DaemonQ, Route, and Infinity. Project neptune. Phrack Magazine, 7(48), 1996.
[17] S. E. Deering. Host extensions for IP multicasting. Internet RFC 1112, 1989.
[18] W. Diffie and M.E. Hellman. New Directions in Cryptography. IEEE TVansactions
on Information Theory, IT-22(6):644-654, Nov 1976.
[19] W. Diffie, P.C. van Gorschot, and M.J. Wiener. Authentication and Authenticated
Key Exchanges. Designs, Codes and Cryptography, 2:107-125, 1992.
[20] DOD. Trusted Computer System Evaluation Criteria. Technical Report DOD
5200.28-STD, Department of Defense, December 1985.
[21] L. Gong. A Security Risk of Depending on Synchronized Clocks. ACM Operating
Systems Review, 26(1), January 1992.
[22] L. Gong and R. Schemers. Implementing Protection Domains in the Java Devel-
opment Kit 1.2. In Proc. of Network and Distributed System Security Symposium
(NDSS), pages 125-134, March 1998.
[23] James Gosling, Bill Joy, and Guy Steele. The Java Language Specification. Addi-
son Wesley, Reading, 1996.
[24] R. Grimm and B. Bershad. Providing policy neutral and transpaient access control
in extensible systems. In Secure Internet Programming [60].
[25] Hermann Hartig, Oliver Kowalski, and Winfried Kiihnhauser. The Birlix security
architecture. Journal of Computer Security, 2(1):5-21, 1993.
[26] C. Hawblitzel, C. Chang, and G. Czajkowski. Implementing Multiple Protection
Domains in Java. In Proc. of the 1998 USENIX Annual Technical Conference,
pages 259-270, June 1998.
[27] M. Hicks, P. Kakkar, J. T. Moore, C. A. Gunter, and S. Nettles. PLAN: A
Programming Language for Active Networks. Technical Report MS-CIS-98-25,
Department of Computer and Information Science, University of Pennsylvania,
February 1998.
[28] Mike W. Hicks and Jonathan T. Moore. PLAN Web Page.
http://www.cis.upenn.edu/'switchware/PLAN/.
[29] C. A. R. Hoare. Communicating Sequential Processes. Communications of the
ACM, 21(8):666-677, August 1978.
[30] C.A.R. Hoare. Communicating Sequential Processes. Prentice-Hall, 1984.
[31] B. Lampson and R. Rivest. Cryptography and Information Security Group Re-
search Project: A Simple Distributed Security Infrastructure. Technical report,
MIT, 1997.
[32] Butler Lampson, Martin Abadi, and Michael Burrows. Authentication in Dis-
tributed Systems: Theory and Practice. ACM Transactions on Computer Systems,
vl0:265-310, November 1992.
[33] X. Leroy and F. Rouaix. Security properties of typed applets. In Secure Internet
Programming [60].
[34] Xavier Leroy. The Caml Special Light System (Release 1.10).
http://pauillac.inria.fr/ocaml.
 450

[35] I. M. Leslie, D. McAuley, R. Black, T. Roscoe, P. Barham, D. Evers, R. Fair-

bairns, and E. Hyden. The Design and Implementation of an Operating System
to Support Distributed Multimedia Applications. IEEE Journal on Selected Areas
in Communications, 14(7):1280-1297, September 1996.
[36] J. Y. Levy, J. K. Ousterhout, and B. B. Welch. The Safe-Tcl Security Model.
In Proc. of the 1998 USENIX Annual Technical Conference, pages 271-282, June
1998.
[37] Francois Louaix. A Web Navigator with Applets in Caml. In Fifth WWW Con-
ference, 1996.
[38] D.D. Clark M.D. Schroeder and J.H. Saltzer. The MULTICS Kernel Design
Project. In Sixth ACM Symposium on Operating Systems Principles, pages 43-56,
1977.
[39] R. Milner, M. Tofte, and R. Harper. The Definition of Standard ML. MIT Press,
1990.
[40] A. B. Montz, D. Mosberger, S. W. O'Malley, L. L. Peterson, T. A. Proebsting, and
J. H. Hartman. Scout: A communications-oriented operating system. Technical
report. Department of Computer Science, University of Arizona, June 1994.
[41] J. Moore. Mobile Code Security Techniques. Technical Report MS-CIS-98-28,
University of Pennsylvania, May 1998.
[42] G. Morrisett, D. Walker, K. Crary, and N. Glew. Prom System F to Typed Assem-
bly Language. In Proc. of the 25th ACM Symposium on Principles of Programming
Languages, January 1998.
[43] Data Encryption Standard, January 1977.
[44] George C. Necula. Proof-Carrying Code. In Proceedings of the 24th Annual
ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages
(POPL), pages 106-119. ACM Press, New York, January 1997.
[45] George C. Necula and Peter Lee. Safe Kernel Extensions Without Run-Time
Checking. In Second Symposium on Operating System Design and Implementation
(OSDI), pages 229-243. Usenix, Seattle, 1996.
[46] Peter G. Neumann. Architectures and Formal Representations for Secure Systems.
Final Report. SRI Project 6401 A002, SRI International, October 1995.
[47] R. De Nicola, G. L. Ferrari, and R. Pugliese. Types as specifications of access
policies. In Secure Internet Programming [60].
[48] Digital Signature Standard, May 1994.
[49] Secure Hash Standard, April 1995. Also known as: 59 Fed Reg 35317 (1994).
[50] Cracker Attack Paralyzes PANIX. RISKS Digest. Volume 18. Issue 45., September
1996.
[51] C. Partridge and A. Jackson. Smart Packets. Technical report, BBN, 1996.
http://www.net-tech.bbn.com-/smtpkts/smtpkts-index.html.
[52] Jon Postel. User Datagram Protocol. Internet RFC 768, 1980.
[53] Jon Postel. Internet Protocol. Internet RFC 791, 1981.
[54] J. H. Saltzer. Protection and the Control of Information Sharing in Multics. In
Communications of the ACM, pages 388-402, July 1974.
[55] M. D. Schroeder. Cooperation of Mutually Suspicious Subsystems in a Computer
Utility. PhD thesis, MIT, September 1972.
[56] M.D. Schroeder. Engineering a Security Kernel for MULTICS. In Fifth Symposium
on Operating Systems Principles, pages 125-132, November 1975.
[57] J. M. Smith, D. J. Farber, C. A. Gunter, S. M Nettles, D. C. Feldmeier, and
W. D. Sincoskie. SwitchWare: Accelerating Network Evolution. Technical Report
MS-CIS-96-38, CIS Dept. University of Pennsylvania, 1996.
 451

[58] P. Syverson. A Taoconomy of Replay Attacks. In Proceedings of the Computer

Security Foundations Workshop VII (CSFW7), June 1994.
[59] J.D. Tygai and Bennet Yee. DYAD: A System for Using Physically Secure Copro-
cessors. Technical Report CMU-CS-91-140R, Carnegie Mellon University, May
1991.
[60] Jan Vitek and Christian Jensen. Secure Internet Programming: Security Issues
for Mobile and Distributed Objects. Lecture Notes in Computer Science. Springer-
Verlag Inc., New York, NY, USA, 1999.
[61] T. von Eicken. J-kernel a capability based operating system for Java. In Secure
Internet Programming [60].
[62] R. Wahbe, S. Lucco, T. E. Anderson, and S. L. Graham. Efficient Software-based
Fault Isolation. In Proc. of the 14th Symposium on Operating System Principles,
pages 203-216, December 1993.
[63] Ian WaJ^eman, Alan Jeffrey, Rory Graves, and Tim Owen. Designing a Pro-
gramming Language for Active Networks, submitted to Hipparch special issue of
Network and ISDN Systems, June 1998. http://www.cogs.susx.ac.uk/projects/-
safetynet/papers/isdn.ps.gz.
[64] David J. Wetherall, John Guttag, and David L. Tennenhouse. Ants: A toolkit
for building and dynamically deploying network protocols. In IEEE OpenArch
Proceedings. IEEE Computer Society Press, Los Alamitos, April 1998.
[65] Bennet Yee. Using Secure Coprocessors. PhD thesis, Carnegie Mellon University,
1994.
[66] P. Zimmerman. PGP User's Manual, 1995.
 Using Interfaces to Specify Access Rights

J. Hulaas, A. Villazon and J. Harms

Abstract. Mobile agents are usually expected to execute in open en-

vironments. This openness implies that they should be able to dynami-
cally learn how to interact with other agents and services which were not
known at development time. The interlocutors therefore have to publish
enough information about their functionality, while at the same time
they have to restrict access rights in order to preserve their integrity.
We describe in this paper a messenger-based framework which proposes
run-time generated interfaces to address this duality.

1 Introduction
A mobile agent is an executing program which can migrate from machine to
machine in a heterogeneous network under its own control. Its utility resides in
the fact that, by enabling the movement of both code and data, it allows the pro-
grammer to reduce network load or to create application-specific communication
protocols. Implementing reliable and secure agent systems is however a difficult
task. The reason is that mobile agents are to execute autonomously in inherently
open and dynamic environments, which moreover may be characterized as:

— Vulnerable: Their openness makes them an easy target for direct attacks.
The security issues in mobile agent systems have been classified into five
categories [1]: (1) transfer security, (2) authentication and authorization, (3)
host system security, (4) computational environment security, and (5) mobile
agent system security. We are interested in the second and especially the
fifth aspects of security, which cover the problems related to initiating and
maintaining secure interaction between mobile agents meeting or residing in
the same computational environment on a given host.
— Completely decentralized: There is no global management of user identities,
and heterogeneity is common in offered services.
In our approach, we do not count on the availability of a unique user identi-
fication, because it does not scale well, and prefer to consider mobile agents
as anonymous entities. By heterogeneity, we refer to the fact that available
services may be implemented in different ways and may offer different in-
teraction protocols from site to site. Mobile agents must therefore have the
abihty to adapt to each situation.

To address these two issues, we propose a single concept, the interlocutor-

specific interface, that is both a view, i.e. a compound capabihty, describing
what operations the actual client is authorized to use, and an extended inter-
face, specifying, e.g. through the associated profiles, how to use the published
 454

operations. The advantage of this solution is that the interlocutor knows noth-
ing more than what is contained in the granted interface. This entails, first, that
the security policy is completely transparent to the client, since the presence
of forbidden operations is never disclosed and always reported as non-existent,
and second, that the process of learning how to interact with an interlocutor is
greatly simplified by the fact that we prevent aquisition of information which
is not relevant. More generally, a different semantics may be attached to each
interface, according to the nature of each interlocutor. In this approach, the
agent or service dynamically generates an interface corresponding to the client's
specific role^.
The idea of having several interfaces is now usual in object-oriented lan-
guages, such as Java, which is e.g. the language of the Aglets mobile agent
system [18]. The first use is to provide a gradation in the levels of visibility on
implementation details: direct instances, heirs and clients are not granted equal
accessibility to the definition of a given class. The other use pertains to the fre-
quent need to have several views of a single implementation, as attested by the
inclusion of the interface construct of Java, which partly replaces the idiomatic
uses of multiple inheritance. For instance, a client object may have to display a
specific view of itself, independently of its true implementation provider, in order
to be acceptable for a service which interacts with its clients through call-backs.
In usual object-oriented languages this is all static, i.e. the number of interfaces
and the contents of these interfaces are defined at compile-time, as well as who
has the right to access them. Interfaces are not obtained by dynamic evaluation
of access rights, they only depend on centralized and implementation-specific
criteria such as class names. Moreover these languages generally do not directly
provide support for building invocations dynamically with the appropriate ar-
guments.
Our initial hypotheses result in a solution which is necessarily quite differ-
ent from more static approaches. Contrarily to other mobile agent systems (e.g.
based on Java[3]), in the messenger paradigm [12], that we have taken as ba-
sis for this work, (a) there is no semantic verification of an agent's code prior
to its execution (as opposed to Java's class loader mechanism), and therefore
we can take advantage of features like dynamic typing and run-time generation
of executable code; (b) there are no predefined code libraries on the platform;
this translates into a generally more adaptative behaviour, but also makes more
difficult the implementation of infrastructure for checking digitally signed code;
(c) there is no predefined equivalent of a direct method call to a mobile agent,
because communication is rather performed through the platform's shared mem-
ory; the agent's protection domain is its private memory space, so that it has
absolute control over the visibility of its data. These are the foundations of our
approach and constitute the specificity of our proposal.
We try to provide a uniform answer to the design issues encountered when de-
signing complex applications relying on secure cooperation between many mobile

' The interface can be seen as a meta-level [24] which performs the access right veri-
fications.
 455

agents from different origins. For instance, unlike objects in statically configured
distributed applications, mobile agents cannot maintain permanent references
to directly locate each other and communicate through access points known in
advance. They need directories to find and to identify each other, as well as a
uniform mechanism to initiate interaction, whether they belong to the same ser-
vice or not. Moreover, mobile agents running under the same authority should
not have the same rights if they implement different functionalities, and mobile
agents implemented by different parties should be viewable as equivalent if they
fulfill the same task. Therefore we promote an approach where each agent's access
rights are determined according to its functionality, and not solely dependent on
its implementation or authority.

Overview of Inter-agent Interaction Mechanisms

Mobile agents do need to synchronize and communicate in order to collaborate to
the accomplishment of higher-level operations. A mobile agent may interact with
other mobile agents (e.g. to delegate the execution of sub-tasks) or with resident
agents which implement local access points to distributed services. There exist
numerous examples of distributed services for mobile agents in the literature:
directory services as in Telescript clouds [15], distributed shared memory and
distributed semaphore [4], resource monitoring [11], resource trading [13] [14],
service brokering and electronic cash validation [17], reliable communication [16].
Hence mobile agents do not only interact as peers, but there also clearly exists
a necessity of layering services in order to build higher-level abstractions.
Inter-agent communication mechanisms may be classified into four categories
[1]: shared memory (as in the Messenger paradigm [11] and sometimes in [3] ),
generative communication (as in the Secure Object Spaces of [1]), datagrams (as
in Agent Tel [14]), and procedure/ method invocations (as in Telescript [15] and
Java). The work described in this paper is based on the Messenger approach,
and extends it to a system based on procedure calls and method invocations.
We show a framework where mobile agent services protect their integrity by
providing interfaces customized to their interlocutors. These interlocutors have
to certify their rights, for instance through a public-key authentication scheme.
Providing a multi-level interface structure is useful as a homogeneous approach
to the management of security in both vertical (i.e. between client agents and
agents furnishing a service) and horizontal (i.e. intra-service) mobile agent inter-
actions. Other existing approaches, which only offer a fixed number of interfaces
(usually two or three), do not allow an agent to modulate in a uniform and fine-
grained way access rights between different kinds of clients (e.g. basic ones and
privileged ones) or between different agents of the same service or organization
(e.g. agents which transfer private data and maintenance agents whose role is to
update the code of resident agents, or to shut down a distributed service in a
coordinated manner). Moreover, interfaces may often not be obtained dynami-
cally, although it is the most natural approach in changing environments. Also,
access rights usually depend on static and language-dependent criteria such as
types or implementation modules. The solution presented here allows anonymous
 456

agents to interact synchronously and asynchronously, without relying on prede-

fined meeting mechanisms, and is implemented efficiently since security checks
need to be performed only once, at the initiation of the interaction.
The rest of this paper is organized as follows. Section 2 presents the Messen-
ger paradigm. Section 3 introduces the mechanisms and details the motivation
for integrating our extended interface structure. Section 4 shows typical situa-
tions where interlocutor-specific interfaces are needed. Section 5 discusses our
approach and compares it with related work. The paper is concluded by Section
6.

2 T h e Messenger P a r a d i g m

In the Messenger paradigm [12], communication between computers is realized

through exchange of programs called messengers. Messengers are anonymous
and autonomous threads of control that are executed on messenger platforms in
a distributed environment, without any form of centralized control. Platforms
are connected by unreliable communication channels, and such an aggregation
of platforms is called a platform system.

messenger execution platform

dictionary
messengers
X -123
y ~ ' abc'

rrr-ri prOCeSS
^-'-'-^ queues
channels

'•. arriving
messengers ^
Fig. 1. Logical view of a messenger platform

Messenger platforms provide directory-like structures, called dictionaries,

which are the fundamental means for organizing storage and exchange of arbi-
trary data between messengers on a same platform. Dictionaries are constituted
as lists of tuples (key, value) and may be protected by Unix-like attributes
rwx telling respectively whether they are readable, writable and browsable. A
 457

dictionary which is not browsable may not be listed: one can then only get the
values for which one knows the corresponding key (Figure 1).
Messengers are basic mobile agent entities which execute in their own local
context defined by a private address space. This constitutes the messengers'
basic protection domain. Messengers do not directly provide any interface or
callable procedures. They can however communicate through a platform-level
shared memory (Figure 2) and coordinate their work by the means of queues.
Data in the shared memory area (e.g. the string ' a b c ' of Figure 2) remains
hidden as long as no reference to it is published by its creator, by insertion in a
global browsable dictionary (the key y is here associated to the string ' a b c ' ) .

private memory private memory

shared memory ^^^^^^ /

X ~ 123
y ~ — _ ^\ ' a b c '
string
o

Fig. 2. Private and shaied memory areas

The messenger platform supports an architecture for building distributed ap-

plications by composition of different services. This architecture makes possible
the secure publication of services which are offered either by single transient
messengers or by organized groups of messengers, which we call families. A fam-
ily is composed of "sedentary messengers" ^ which fulfill a service locally and
of helper messengers which migrate from platform to platform to convey infor-
mation or to perform other auxiliary tasks on behalf of the sedentary ones. This
corresponds to a vision of the network where agent services tend to aggressively
people platforms as they become available.

^ In principle a resident or sedentary messenger will not migrate after its arrival on
a given platform, unless it is forced to (local resources too expensive, not enough
clients, detection of threats to its security, etc.)
 458

2.1 Service Publication Mechanism

Service publication is performed by creating in the shared memory area a dic-

tionary which then functions as platform-level interface to the service. This data
structure is called the Service Access Point (SAP). All interfaces are addition-
ally described using the Messenger Interface Description Language (M-IDL) [4],
which must be understood by the clients. M-IDL essentially is a grammar which
specifies how to use the service through procedure profiles and rules for com-
posing correct sequences of procedure calls. This already gives a good idea on
how to use the service {MIDL does however not give a way to specify semantics).
M-IDL specifications are intended to be interpreted at run-time; we can consider
them as executable specifications.

shared memory of a platform

servdict (r\w[x)
secretdef distsema"

globaldict (r{w)
-iv seinadict (r\x) (service access point)
c ef k2
fullname ~ "distributed semaphore"
version - "1.0"
P - procedure (x)
V - procedure(x)
create ~ queue

Fig. 3. Single-level secure service publication architecture

A messenger wishing to publish its service will choose a secret key kl and
generate a public key k2 (using a one-way function). The tuple (k2, SAP) is
inserted into the global dictionary globaldict. Then the service provider has to
add to the service dictionary servdict an entry (k2, e x p l i c i t s e r v i c e name).
Because s e r v d i c t is a browsable dictionary, any messenger can search for the
name of the service it is looking for; however only messengers possessing the
original key kl can remove the corresponding entry.
Clients will gain access to a service by the means of its name (e.g. distsema as
in fig. 3). They will then use the corresponding key k2 to obtain the desired SAP
from globaldict (which is not browsable). The latter operation will finally enable
the client-furnisher interaction: in the SAP, each procedure name is associated
 459

with the corresponding code, which is not readable (and thus may not be copied),
but only executable.
This first messenger service publication mechanism uses a uniform interface
structure allowing for elementary interactions between messengers. In this model,
a unique access policy is enforced for all service clients, with a single level of
protection and no possibihty of discrimination. Moreover, messengers of a same
family must access internally shared information through ad-hoc strategies, in-
volving e.g. predefined knowledge of hidden data structures, instead of resorting
to homogeneous protected interfacing mechanisms. For higher-level abstractions,
such as messenger families, more elaborated schemes providing multiple protec-
tion levels are needed. Our solution is described in the following section.

3 Interlocutor-Specific Interfaces

As an extension to the previously described publication mechanism, we designed

a generic pattern for intra- and inter-family protection and elaborated a corre-
sponding multi-level interface model. This model enables services to apply finer
and more efficient protection schemes; we use it for instance to implement the
visibility rules commonly found in object-oriented languages (public, protected
and private). In this approach, the service provider associates a customized in-
terface to each client, i.e. the architecture allows the client to obtain the interface
corresponding to its specific rights.

, . . ,. Semaphore
® obtam mtertace service

name ~ "distributed semaphore"

k2 - "distsema" k2 - version ~ "2.0"
Getlnterface{) ~

Fig. 4. Obtaining an interface

 460

We have seen that the only criteria for accessing a service SAP is the knowl-
edge of the service name. Phases 1 and 2 of Figure 4 correspond to the single-
level interface access to SAP. Phases 3, 4, 5 and 6 are related to the more
elaborate protection scheme. Instead of obtaining directly a complete SAP, the
G e t l n t e r f a c e O procedure which is found in phase 3 must be called by the
client, who also gives as argument a secret key for authentication. On the basis
of this authentication, the service will decide to which level it wants to grant
access to its own functionality, and return to the client a correspondingly gen-
erated interface. Note that the simplistic authentication scheme outlined here is
just a conceptual view; in reality we may need symmetric keys or some ad-hoc
infrastructure such as Kerberos [5].
The secret key described here determines how a client wants to identify itself.
Several identities may therefore be adopted: this may be useful if a messenger is
a composition of two services and must be viewable alternatively as belonging to
the one or to the other. A family is thus characterized by the possession of a se-
cret key which enables each messenger to authenticate itself w.r.t. other members
of the family. For security reasons, a messenger's membership of one (or several)
families is not a visible attribute. Similarly, messengers do not have identities
(they are anonymous) and cannot be directly referenced either ^. A family is
not the same as a type, because small auxiliary messengers, which receive lim-
ited functionality for executing precise sub-tasks, such as to spread information
within a distributed service, are also members of the family implementing this
service.
The granted interface, with its (procedure name, procedure code) tuple
set, is created in the receiver's private memory. The code associated to each pro-
cedure name is as usual accessible in execute-only mode, which prevents inspec-
tion or further copying (i.e. migration or transmission) of sensitive information
to untrusted parties. This very code is just a set of stubs which hide the location
of the actual code and data in the global memory. The interface is not trans-
ferable and is therefore equivalent to a session key. It could also be viewed as a
proxy for an object which is not remote, but simply in another local protection
domain. It must be noted that it is harder to implement non-transferable objects
in systems such as Java, which do not provide separate address spaces within
a single virtual machine. One possibility is to let the interface stubs check that
the current client is the one that originally received the interface, which means
that every agent must have an accessible unique identificator. The paper [25]
describes the difficulties in implementing Java-based capabilities, a notion which
is close to our interfaces.

4 Uses of Interlocutor-Specific Interfaces

The protection scheme we describe in this paper is usable every time an agent
has to interact with many other kinds of agents and therefore wants to only
Compare with Java's synchronized statement which allows anyone to freeze an ob-
ject through a reference
 461

partially reveal its functionality. Let us first consider a banking example. A

bank has many kinds of clients: basic, private customers, VIPs, or international
corporations, just to mention a few. Each category has its own set of burdens
and privileges. The available operations are not the same (the international
corporation may e.g. additionally discuss interest rates) and their usage are also
specific (the VIP is not asked to check the balance before withdrawing money
from a given account). The bank will therefore disclose to each client only those
operations which are available to him, and provide together the simplest possible
description. If you now consider the employees of the bank, the situation is the
same, although they all belong to the same service and should have some kind of
"private" access rights. They have different duties and prerogatives, according to
their hierarchical rank and to which department they work in. Additionally, the
rights of the basic employee, when he is acting as a client, do not even enclose
those of the most important clients, as would be the case in a classical computer
language.
We outline below two generic situations where one can make profitable use
of the interlocutor-specific interfaces. The first we present is, in the frame of our
notion of messenger family, the Resident-Emissary pattern. This description is
performed following the model of [19]. The second use is demonstrated by a new
object-oriented dialect of the M 0 messenger programming language, and the
protection mechanisms provided by the associated compiler.

4.1 The Resident-Emissary Pattern

Intent: The Resident-Emissary pattern allows collaboration between stationary
agents over multiple hosts, as well as any form of intra-service critical interaction.

Motivation: When implementing distributed mobile agent services (i.e. dis-

tributed applications implemented with mobile agents), one can imagine dis-
tribution and installation phases in order to provide local service publication
and support over multiple platforms (these phases are possible after the discov-
ery of newly available resources for example). The local support is implemented
by agents that remain on the platform to coordinate the service's distributed
functionality. We qualify such agents as sedentary or resident.
The real functionality of the distributed service requires communication be-
tween resident agents. This might be e.g. to multicast new information, as when
a member of an indexing service wants to propagate a new service or resource
to be promoted.
The information which is exchanged should not be disclosed. Therefore a
carrying agent, that we call the emissary agent, must have the necessary privi-
leges to access the resident agents' updating input channels. These channels are
visible only through a particular interface, the family's private interface.
It is preferable to use emissary agents, rather than resorting to direct in-
teraction between residents, because the emissary contributes to the protection
of sensitive information (stored in the family's hidden local repositories), by
accepting to access the repositories only by the means of tailored interfaces.
 462

Applicability: Use this pattern when:

- Distributed entities must interact to exchange data which ensure the service's
coherency.
- Special "signals" between resident agents (that imply radical changes of be-
havior) must be delivered.

Participants:
Resident: Defines the stationary agent that performs the local service imple-
mentation and that has to exchange critical information with other station-
ary agents.
Emissary: Mobile agent which transports the information between resident
agents.
Interface: Means by which selective access to vulnerable information is pro-
vided.
Repository: Location of information that must be only accessible through spe-
cific interfaces.

Collaboration: The Resident needs to communicate private data with other

residents of the same family.
- On the local platform (say Platform A) Resident A creates an emissary
agent with a specific itinerary in order to contact the remote resident (say
on Platform B).
- The Emissary contacts its creator, asking for an interface that indirectly
allows him to access data stored in the local repository (note that no direct
access is allowed, all requests are processed through the interface).
- On arrival, the Emissary agent performs a similar interaction, contacting
Resident B on the destination platform and requesting an interface.
- Resident B will provide a tailored interface according to its rights (allowing
access to the family's update routines).
- Finally the Emissary will be able to update the information through the
corresponding interface.
Tailored interfaces are only usable by the original requesting agent, and they
are automatically discarded when this agent migrates or terminates. Emissary
agents do not necessarily return to their origin.

Consequences: The Resident-Emissary pattern is useful for the design of fam-

ilies which need secure communication of critical data. The usage of emissary
agents also eases, for example, incremental collection and update of information
inside a service residing on several platforms. Indeed, an emissary can update its
data each time it contacts another resident agent. Furthermore, even its itinerary
can be dynamically adapted to new destinations (stored by residents as private
data) leading to a globally more flexible and eflicient approach than remote
procedure calls.
 463

Platform A Platform B

ResWent Repository Prjvate_ ^^^^^„^ ^ . . Private Repository Resident

Emissary ,„,g^3j.g B B
Interface
creates

getlnterface(I

creates

getData{ initO

getlnterface(I

creates
update{)

Fig. 5. Collaboration in the Resident-Emissary pattern. Agent migration is represented

as dashed links with centered arrows. Names suffixed with a star, e.g. AnAgent*, denote
agents after migration (i.e. arrived on a remote platform).

4.2 The Object-Oriented Compiler

Another situation where multiple interfaces and access levels are necessary is a
messenger on-the-fly compiler service which makes possible the implementation
of object-oriented systems. We have defined an object-oriented dialect for our
messenger programming language M 0 . The idea is then to allow messengers to
carry with them the compiler for this dialect as well as the code of the programs
to be compiled, so a messenger family can exploit its own class definitions for
the collaboration between family members. A messenger will install the compiler
code in the local platform's service dictionary making it available as a public
utility (Step 1 in Figure 6). Other messengers (belonging or not to the same
family) may then use it in order to compile and install new messenger services
(Step 2). Finally, clients may perform requests resulting in the creation in their
local memory of instances of those classes (Steps 3 and 4).
In the new object-oriented dialect we find the usual notions of class, method,
object, instance variable, and single inheritance. Here we see the natural need for
multi-level visibility, with messengers applying a three-level visibility to grant
access according to family needs. Thus, members of the family have access to
private data, while only public data is visible to clients. The protected keyword
denotes a special inter-family mode, based on local criteria. In Java, the pro-
tected notion denotes visibility given to classes in the same 'package'. The idea
that we applied in our approach is that a messenger which wants to derive a sub-
class from a given hbrary class, must obtain the right to do so. This is necessary
because a class must expose part of its own internal structure to enable inheri-
 464

mO cod'

Compiler
Classes
shared memory

Fig. 6. The compiler as support for public services implemented as classes

tance. Practically, this is realized by adding a class method called inherit, which
is available in its full form only through the protected interface of the class. Mes-
sengers which do not have enough privileges will only obtain a crippled version
of the inherit method through the public interface, and thus they will not be
able to inherit entities defined with the protected visibility mode. This approach
partly solves the security holes of Java related to malicious sub-classing as de-
scribed in [1]: now, if a client trusts a given root class, he can be more confident
that the actual instances he deals with are not harmful to him, and that he will
not receive harmful objects profiting from the sub-typing mechanism.

Another noteworthy characteristic of this compiler is that all instances gen-

erated by the classes it has compiled are systematically authenticated before any
method call as part of the dynamic binding process. This is made possible by
the insertion of a secret pattern (a kind of "magic cookie ") inside every instance
created. This part of the instance cannot be explored by anyone, even the client,
although it resides in a private memory area, since it is implemented as a non-
browsable dictionary. If the instance must migrate, it must be authenticated and
unmarshalled by a corresponding class on the destination platform in order to
re-insert a new secret pattern. It is therefore not possible for malicious clients to
forge instances which threaten the integrity of the service implemented by the
class. This approach is reminiscent of Telescript's notion of telename, which is
a unique and nonforgeable agent (or place) identifier; since a telename identifies
a process, or a set of processes, it does however not provide the fine-grained,
class-level protection described here.
 465

5 Related Work
Some distributed systems include a notion of computational reflection which
enables them to inspect and modify their own structure, in particular their
interface. The HADAS system [21] for instance has mobile objects, called am-
bassadors, which provide an interface for remote applications and can be seen as
an advanced form of proxy. Ambassadors are reflexive in the sense that they can
change their interface according to local conditions, and can receive and even
tailor part of the functionality of the remote application they represent. They
however are not autonomous, since they cannot migrate any further, which is
similar to the limitation of Java applets [3]. They do not either exploit their
computational reflection with the objective of enhancing security.
In Dalang [20], reflection and code modification are used to adapt software
components as they are downloaded into the executing environement. This work
is based on Java and is not specific to the mobile agent domain. When the byte
code is loaded, a specialized class loader analyzes it and modifies the behaviour
by generating an interface with code to catch methods calls. This approach is
thus based on a meta-object protocol that transparently enables security-related
verifications. Interface generation in the sense of Dalang does not convey the
same meaning as in our approach, because the single interface that is created
is shared by all interlocutors. Another difference is that the associated code
generation mechanism is not triggered at run-time, but at load-time.
An interface in the Distributed Component Object Model (DCOM) [6] is a
set of functions bound to a certain object which implements them. Each object
may introduce several interfaces and a user may query one of them using the
Query I n t e r f a c e function, which itself belongs to a default interface supported
by every object. Query I n t e r f a c e changes the semantics of the object as seen by
its user. This approach is limited to traditional distributed systems and does not
address mobile agents. Authentication is not performed by DCOM itself, but by
the lower layers it is built upon (DCE RPC [7]). It is now unclear how well these
two levels integrate, and whether the delivered interfaces can be customized on
the basis of the results of the authentication in order to enhance security.
In the frame of CORBA [8], several current initiatives tend to answer is-
sues described in this paper. The Multiple Interfaces RFP [9] deals with the
resolution of conflicts between multiple interfaces on the same object, while the
Composition Facility [9] provides the means for objects to be composed of log-
ically distinct services by the use of multiple interface definitions. In CORBA
in general, as well as in its draft Mobile Agent System Interoperability Facility
(MASIF) [10], authentication for method invocations is performed on a call-by-
call basis, and interface delivery is not under the control of the implementing
objects, but of the Interface Repository mechanism.
In our approach, interfaces are not designed to be transmitted for remote
usage. The stub routines they contain cannot be migrated because the corre-
sponding code is execute-only and may not be copied. This is consistent with
the mobile agent philosophy, which tries to limit the frequency of remote method
invocations, and is indissociable from our security architecture. The granting of
 466

a dedicated interface is equivalent to a session key, which is more efficient than

authentication on a call-by-call basis, and more secure than the publication of a
single, big, public interface. Revocation of an interface by its creator is possible:
the stub routines can check whether they should continue to service method calls,
and if not, they can destroy themselves in order to eliminate any unwanted ref-
erences. Any invocation will from there on result in an exception on the caller's
side.
To our knowledge, the present paper presents the first approach to integrate
the use and support for multiple interfaces in the mobile agent world. As already
noted in [22], agents move towards the services they need, and then adapt to
the interface they obtain. This enables a more flexible and efficient RISC-like
interaction, as opposed to coarse-grained CISC-like remote method invocation
protocols. This remark also strengthens the motivation for supporting multiple
specialized interfaces.
This interaction mechanism is described for more clarity as happening be-
tween a service (implemented by a/resident messenger) and an incoming messen-
ger. It can however be generalized to the synchronization and communication
between any pair of agents hosted by the same platform; to this end, they should
locate each other not through the service dictionary, but through any other global
dictionary.
Finally, we do not address in this work the problem of protecting messengers
from malicious hosts. This issue remains largely unsolved by current research,
although some encouraging results have been found lately [23]. In this paper
we are only interested in inter-agent relationships, which makes possible the
reliance on platforn^-provided mechanisms such as separate address spaces and
the protection of code by an execute-only access mode.

6 Conclusion

We described in this paper a framework where anonymous mobile agents protect

their integrity by providing interfaces customized to their interlocutors. By ex-
ploiting the characteristics of the Messenger platform, such as separate address
spaces and execute-only code segments, it is possible for an agent to disclose only
the information which is needed for other agents to securely interact with him.
We showed that this customer-specific interface structure is useful in both verti-
cal (i.e. client-furnisher) and horizontal (i.e. intra-service) relationships. It offers
a level of flexibility and security which is unavailable in existing object-oriented
or type-based approaches, but strongly needed for a wide deployment of mobile
agents in open and dynamic environments. Our research was performed in the
frame of the Messenger paradigm, which implies a rather low-level approach. In
other words, we cannot force the programmer to write secure programs, but we
give him the tools to ease his effort in that direction.
 467

Acknowledgements
T h e authors would hke to t h a n k the anonymous reviewers for their comments
and useful suggestions. This work was funded by t h e Swiss National Science
Foundation grant 20-47162.96.

References
1. J. Vitek, M. Serrano and D. Thanos, Security and Communication in Mobile Ob-
ject Systems, in Mobile Object Systems: Towards the Programmable Internet, Sec-
ond International Workshop, MOS'96, Linz, Austria, Selected Presentations and
Invited Papers, J. Vitek and C. Tschudin (eds), LNCS vol. 1222, July 1996.
2. C. F. Tschudin, An Introduction to the M 0 Messenger Language, Technical Report
86, (Cahier du GUI), University of Geneva, 1994.
3. J. Gosling and H. McGilton. The Java Language Environment. A White Paper.
Sun Microsystems, May 1995.
4. M. Muhugusa. Distributed Services in a Messenger Environment: The Case of
Distributed Shared-Memory. Ph.D. Thesis no 2903, University of Geneva, 1997.
5. J. G. Steiner, B. Clifford Neuman, and J.I. Schiller, Kerberos: An Authentication
Service for Open Network Systems, In Proceedings of the Winter 1988 Usenix
Conference, February 1988.
6. N. Brown and C. Kindel, Distributed Component Object
Model Protocol - DCOM/1.0, Internet draft, January 1998,
http://www.microsoft.com/oledev/olecom/draft-brown-dcom-vl-spec-02.txt
7. CAE Specification, X/Open DCE: Remote Procedure Call, X/Open Company Lim-
ited, X/Open Document Number C309. ISBN 1-85912-041-5, Reading, Berkshire,
UK, 1994.
8. Object Management Group, The Common Object Request Broker: Architecture
and Specification (Revision 2.0), Object Management Group, Framingham, Mass.,
1995.
9. Object Management Group, Multiple Interfaces and Compo-
sition, work in progress, April 1998, Web information page:
http://www.omg.org/library/schedule/Multiple_Interfaces.and_Composi-
tion.htm
10. Object Management Group, The Mobile Agents Facil-
ity, work in progress, April 1998, Web information page:
h t t p : //www. omg. org/library/schedule/Hobile_Agents_Facility_RFP. htm
11. A. Acharya, M. Ranganathan, J. Saltz, Sumatra: A Language for Resource-Aware
Mobile Programs, in Mobile Object Systems: Towards the Programmable Internet,
Second International Workshop, MOS'96, Linz, Austria, Selected Presentations
and Invited Papers, J. Vitek and C. Tschudin (eds), LNCS vol. 1222, July 1996.
12. C. F. Tschudin, The Messenger Environment M 0 - A Condensed Description, in
Mobile Object Systems: Towards the Programmable Internet, Second International
Workshop, MOS'96, Linz, Austria, Selected Presentations and Invited Papers, J.
Vitek and C. Tschudin (eds), LNCS vol. 1222, July 1996.
13. C. F. Tschudin, Open Resource Allocation, First International Workshop on Mobile
Agents (MA'97), Berlin, Germany, April 1997.
14. R. Gray, Agent TCL: A Flexible and Secure Mobile-Agent System, in Proceedings
of the fourth annual Tcl/Tk Workshop (TCL 96), July 1996.
 468

15. J. E. White, Telescript Technology: The Foundation for the Electronic Market-
place, General Magic White Paper, General Magic, Inc., 1994.
16. C. F. Tschudin, On the Structuring of Computer Communications, Ph.D. Thesis,
University of Geneva, Switzerland, 1993.
17. D. Johansen, R. van Renesse and F. B. Schneider, Operating System Support for
Mobile Agents, in Proceedings of the 5th IEEE Workshop on Hot Topics in Op-
erating Systems, pages 42-45, Orcas Island, Wash., May 1994. Also available as
Technical Report TR94-1468, Department of Computer Science, Cornell Univer-
sity.
18. D. Lange, M. Oshima, G. Karjoth and K. Kosaka, Aglets: Programming Mobile
Agents in Java, in 1st International Conference on Worldvifide Computing and
its Applications (WWCA'97), T. Masuda, Y, Masunaga and M. Tsukamoto, Eds,
LNCS vol. 1274, Springer, Berlin, Germany, pp. 253-266, 1997.
19. Y. Aridor, D. Lange, Agent Design Patterns: Elements of Agent Application De-
sign. Second International Conference on Autonomous Agents (Agents'98). Min-
neapolis/St. Paul, May 10-13, 1998.
20. Ian Welch and Robert Stroud, Dynamic Adaptation of the Security Properties of
Applications and Components., ECOOP Workshop on Distributed Object Security,
Brussels, Belgium, July 1998.
21. O. Holder and I. Ben-Shaul, A Reflective Model of Mobile Software Objects, in
Proceedings of the 17th IEEE International Conference on Distributed Computing
Systems (ICDCS'97), Baltimore, Maryland, USA, May 27-30 1997.
22. F. B. Schneider, Towards Fault-tolerant and Secure Agentry, Invited paper, 11th
International Workshop on Distributed Algorithms, Saarbrcken, Germany, Sept.
1997.
23. T. Sander and C. F. Tschudin, Towards Mobile Cryptography, In proceedings of
Security&Privacy'98, May, 1998.
24. G. Kiczales, J. des Rivieres and D. G. Bobrow, The Art of the Metaobject Protocol,
MIT Press, 1991.
25. T. von Eicken, J-kernel a capability based operating system for Java, In Secure
Internet Programming, Lecture Notes in Computer Science, Springer-Verlag Inc.,
New York, NY, USA, 1999.
Introducing Trusted Third Parties to the Mobile
Agent Paradigm

Uwe G. Wilhelm, Sebastian Staamann, and Levente Buttyan

Abstract. The mobile agent paradigm gains ever more acceptance for
the creation of distributed applications, particularly in the domain of
electronic commerce. In such applications, a mobile agent roams the
global Internet in search of services for its owner. One of the problems
with this approach is that malicious service providers on the agent's
itinerary can access confidential information contained in the agent or
tamper with the agent.
In this article we identify trust as a major issue in this context and
propose a pessimistic approach to trust that tries to prevent malicious
behaviour rather than correcting it. The approach relies on a trusted
and tamper-resistant hardware device that provides the mobile agent
with the means to protect itself. Finally, we show that the approach is
not limited to protecting the mobile agents of a user but can also be
extended to protect the mobile agents of a trusted third party in order
to take full advantage of the mobile agent paradigm.

1 Introduction
New approaches for distributed computing based on mobile agent technology,
such as Aglets, Telescript, or Voyager become ever more pervasive and are con-
sidered as innovative new ideas to structure distributed applications. A partic-
ularly interesting and perhaps economically important class of applications, to
which mobile agents seem well adapted, is electronic commerce.
A typical use of mobile agents in the domain of electronic commerce includes
the scenario, in which an agent roams the global Internet in search of some service
for a user (the owner of the agent). Such a service can have many different forms,
for instance, the provision of a physical good, the execution of a search for an
information item, or the notification of the occurrence of some event. The agent
is configured by the user with all the relevant information about the desired
service, the constraints that define under which conditions an offer from a service
provider is acceptable, and a list of some potential providers of the service. It
will then migrate to the sites of these service providers in order to locate the
best offer for the service sought by the user and finalize the transaction with the
chosen service provider.
Since an agent is vulnerable when it is executing on the execution platform
of a service provider, it is necessary that the user obtains some guarantees con-
cerning the protection of his agents. Consider a mobile agent that holds data for
one or several payment methods, which it needs to finalize a purchase. These
 470

payment data should not be available to any principal other than the one that
actually provides the service and, thus, is entitled to receive the payment. A
malicious service provider might try to obtain the data of the payment method
without providing the service or might otherwise tamper with the agent in order
to trick it into accepting the malicious provider's offer (e.g., by removing some
information about a better offer from the memory of the agent). The usual ap-
proach that is taken to provide a user with certain guarantees concerning the
protection of his agents, is to assume that the service providers are trusted prin-
cipals [13] or to create a mechanism that enables the user to detect which of the
providers on the itinerary have misbehaved [21].
The notion of trust has long been recognized as being of paramount impor-
tance for the development of secure systems [6,10,28]. For instance, any con-
ceivable system for authenticating users needs trusted functionality that holds
the necessary authentication information (see e.g., [18,26]). Yet, the meaning
that is associated with trust or the notion of a trusted principal is hardly ever
clearly defined in these approaches and the reader is left with his intuition.
In this article we address the question of how trust in a certain principal can
be motivated based on technical reasoning and present a pessimistic approach
to trust that tries to prevent malicious behaviour rather than correcting it after
it has occurred. The approach relies on a trusted and tamper-resistant hardware
device that can be used to enforce a policy. If this policy is properly chosen, an
agent can take advantage of it in order to protect itself as well as the information
it contains from possibly malicious service providers.
The mobile agent paradigm usually identifies two interacting principals: the
owner of the agent who configures it and the executor of the agent, which may
be identical to the service provider. However, many protocols for security related
problems, especially those that are concerned with non-repudiation, require an
additional third party that often has to be trusted by the other parties. This
principal is called a trusted third party (TTP). Due to this trust requirement, the
functionality of a TTP must be realized in a trustworthy environment, which
is usually only available at the site of the TTP. Hence, the principals in the
mobile agent paradigm have to interact with the TTP server using the classical
client/server mechanisms. The approach described here, which provides protec-
tion for the mobile agents of regular users, can also be applied to mobile agents
that are owned by a TTP. This allows us to take full advantage of the mo-
bile agent paradigm and removes the need for remote messaging between the
interacting principals.
In the following Section 2, we introduce our model for mobile agents and
point out the problems related to trust within this model. Then, in Section 3,
we discuss the notion of trust and define its relation to policy, which enables us
to better assess the possible motivations for trust. In Section 4, we introduce
a trusted and tamper-resistant hardware device and a protocol, which allow us
to define certain guarantees for the execution of agents. In Section 5, we show
how this approach can be used to protect the agents of a regular user as well as
those of a TTP. In Section 6, we discuss why we consider this to be an adequate
 471

way to approach the problem and what effects this has on the notion of open
systems. Finally, Section 7 concludes the paper with a summary of the main
contributions.

2 T h e mobile agent paradigm

The mobile agent paradigm has been identified by many authors as a promising
and innovative approach to structure problems in distributed computing [3,4,7,
8,11,23]. However, it is still under hvely discussion and it has been shown in [9]
that there is no single compelling reason to favour the mobile agent paradigm
over classic client/server approaches. On the other hand, the same authors point
out that the mobile agent paradigm provides interesting solutions to many real-
life problems, for instance in the context of:
— mobile users, where an agent is sent out from a mobile computer in order to
accomplish a well-defined task on behalf of the user while he is disconnected
from the communication network. Once the user reconnects, the agent re-
turns and reports the result of the task or the problems it encountered.
— high-bandwidth interactions, where an agent is sent to a database server
that holds a large amount of unstructured data to search for some specific
information for the user.
— customizable services, where a service provider can offer a small number of
simple operations on its service interface that can be combined by an agent
to obtain the desired functionality.
— resident agents, which are stationary agents that take residence at some
service provider and handle simple routine actions for their owner (e.g.,
communication management for a mobile user, where the agent decides how
to handle an incoming communication request).

2.1 Basic definitions

In this article, we do not focus on the underlying technology that is used to
implement the paradigm, but we only require a simple model of agents for our
discussion. Therefore, we identify the following major abstractions that we asso-
ciate with mobile agents. A mobile agent consists of code, data, and its current
execution state, which can be marshalled by the agent owner in a transport
format and subsequently sent to the agent executor. The agent can be confiden-
tiality and integrity protected during transit to prevent outside attacks through
the use of cryptographic mechanisms. These mechanisms can also provide data
origin authentication for the marshalled agent. The agent executor will then
eventually unmarshall the agent and instantiate it on a special environment
located at the agent executor, which is called agent platform (AP). Here, the
mobile agent can interact with services of the local AP as well as other agents
located at this AP and continue to accomplish the task it was given by its owner.
The literature on agents (e.g., [3,13]) distinguishes between two different
approaches to support agent mobility: weak mobility and strong mobility. The
 472

former does not automatically support the transfer of the current execution
state of the agent. Thus, if an agent is supposed to visit more than a single AP
(which is often referred to as multi-hop agent), the current execution state has
to be explicitly encoded in the agent's data before it can migrate to another AP.
The latter approach automatically supports the transfer of the current execution
state and allows an agent to continue its execution exactly where it left off before
initiating the migration. A mobile agent can thus easily visit as many APs as it
deems necessary to accomplish the desired task. In both approaches, the result
of the agent's remote execution can be sent directly to the agent owner in the
form of a message or kept in the execution state of the agent and extracted by
the agent owner when the agent returns.
The reason for only providing weak instead of strong mobility is that the
execution environment in the AP and the agent transport format can be much
simpler since the AP does not have to provide the current execution state (which
is, for instance, not available from the Java virtual machine) and the transport
format does not have to encode it. Also, if an agent visits only a single AP,
which is supported by weak mobility, the trust model becomes much simpler.
Any damage incurred by the agent to the AP (and thus to the agent executor) or
by the AP to the agent (and thus to the agent owner) can easily be attributed
to the other entity. If more than two principals are involved, the problem of
accountability becomes much more difficult [13]. Each principal can defer any
damage to actions by one of the other principals on the agent's itinerary.
In this article we will concentrate on the protection of a mobile agent on a
particular AP. In our solution, a multi-hop agent that visits several sites only
represents multiple instances of the same problem. Our main concern is how to
protect the agent and especially the data it contains from undue manipulation
by or undesired disclosure to the agent executor, which is mainly a question of
trust in the agent executor.

2.2 The need for trust

There are many examples where an agent might need confidential information
that should not be disclosed to the service provider, even though the agent needs
the information to accomplish its task:
- a shopping agent in an electronic-commerce system might hold data that
could give a bargaining advantage to the service provider if it were known
to him (e.g., a maximum price that the service user is willing to pay or the
lowest QoS that he is willing to accept before inquiring at another service
provider).
- another agent for electronic commerce might hold a private key with which
it can sign messages on behalf of its owner, for instance, to confirm an order
placed by the agent. This key has to be kept secret to prevent that the service
provider can also sign messages on behalf of the agent owner.
- an agent in a personalized information system might contain private data
from the customization by its owner that is needed to find appropriate infor-
mation. An agent searching for movies that are likely to interest its owner.
 473

might contain some very personal information about the user's special in-
terests, which the agent executor cannot infer from simply observing the
agent's choice.
finally, an agent that merely searches for some particular financial informa-
tion (e.g., stock quotes) might, depending on the owner of the agent, convey
some very sensitive information (the mere request already conveys the inter-
est in the information).

2.3 Threats to a mobile agent

In a conventional mobile agent system, when the agent owner sends a mobile
agent to an agent executor in order to use some service, the agent owner loses
all control over the code and data of the agent. The agent executor can:

— reverse engineer the agent's code,

— analyze the agent's data,
— arbitrarily change the agent's code and data, or, if none of these direct
attacks is feasible,
— experiment with the agent (e.g., by feeding it with arbitrary data and reset-
ting it to its initial state, in order to observe the agent's reactions).

This constellation puts the agent executor in a much stronger position than
the agent owner. The agent owner simply has to trust the agent executor not to
use the methods described above to illicitly obtain confidential information from
the agent that it has to carry in order to use the service. There is no way for the
agent owner to control or even know about the behaviour of the agent executor.
The reason for the imbalance between agent executor and agent owner in the
mobile agent model as compared to the principals in the client/server model is
that in the former approach, the agent owner has no guarantees whatsoever con-
cerning the execution of its agent. In the client/server approach, the client relies
on many guarantees that are so basic that one hardly ever thinks of them. Never-
theless, these guarantees allow the implementation of certain types of behaviour
in the client part of the distributed application that can not be implemented
in conventional agent systems (e.g., code will be executed at most once, code
will be executed correctly, or the code can rely on a reasonably accurate time
service). This is due to the fact that the client implementation is under the phys-
ical control of the service user, who can can observe what is happening in the
system and notice any irregularities. Thus, he is able to react accordingly, for
instance, to interrupt an ongoing transaction or to log any irregularities at the
client side so that they can be provided as evidence in the case of a dispute with
some server. This is opposed to the mobile agent paradigm, where logged data
can easily be deleted by the agent executor.
We intend to create an environment for mobile agents that allows them to
base their execution on assumptions similar to the client/server approach, so that
it becomes possible for a mobile agent to better protect itself from a malicious
agent executor.
 474

3 The notion of trust

We already mentioned the importance of trust for security in distributed systems
and pointed out the lack of a clear definition of what is meant by the terms trust
or trusted principal. In the following, we present our analysis of possible trust
relations between different principals.
A reason for the lack of a clear definition of trust could be that trust is
more a social than a technical issue and consequently quite difficult to tackle
entirely in a technical approach. The major problem stems from the fact that
the notion of trust mixes the goals of a principal with its behaviour to achieve
these goals. In order to trust some principal, it is usually necessary to concur
with or at least to approve of its goals (which are not always clearly stated)
and to believe that it will behave accordingly. In our definition of trust, we will
try to clearly separate these two issues by identifying a policy that is consistent
with the goals of the principal. This policy is a set of rules that constrains the
behaviour of the principal for all conceivable situations. It has to be written
down and made available to all other principals that interact with the issuer of
the policy. Then, we define trust in another principal as the belief that it will
adhere to its published policy.
The question of whether a certain principal can be trusted now consists of
(a) checking its published policy in order to decide if it is acceptable and (b) to
establish a motivation for the belief that it will adhere to its published policy.
The former is quite difficult but can be supported by a formal specification
of the security policy (similar to the approach in [15]). The latter, however,
is a problem that is quite difficult to formalize. Depending on how the belief
in the adherence to the published policy is motivated, we have identified two
fundamentally different approaches to the problem of trust:

— the optimistic approach and

- the pessimistic approach.

In the optimistic approach, we give an entity the benefit of the doubt, assume
that it will behave properly, and try to punish any violation of the published
policy afterwards. In the pessimistic approach we try to prevent any violation of
the published policy in advance by effectively constraining the possible actions
of a principal to those conforming to the policy. Both of these approaches have
advantages and disadvantages.

3.1 The optimistic approach

This approach is easy to implement, since it does not require any special mea-
sures to make trusted interaction possible. This is probably the reason why it
is the basis for most business conducted today. On the other hand, it requires
some reliable mechanism to discover a policy violation after it has occurred. If
such a mechanism does not exist, then the approach degenerates to blind trust,
which indicates that there is no particular motivation to believe that a principal
 475

will adhere to its published policy other than its own assertion. Blind trust is
obviously a very weak foundation for trust and not recommended for any im-
portant or financially valuable transaction. It is therefore important to make the
probability that a policy violation is discovered as high as possible by improving
controls and establishing checkpoints.
Once a policy violation is discovered and if it can further irrefutably be
attributed to one of the participants in the corresponding transaction, this prin-
cipal should be punished according to the appropriate laws and the damage
caused by the policy violation. The primary goal of this punishment is to deter
potential violators from committing a policy violation in the first place.
Depending on how this punishment is enacted, we identify the following two
motivations for the belief that an entity will adhere to its published policy:
— trust based on (a good) reputation and
- trust based on explicit punishment.
Trust based on reputation stems from the fact that the principal in question
is well known and has very little to gain through a violation of its own policy
but a lot to lose in case a policy violation is discovered. This loss is supposed
to transpire from the lost revenue due to customers taking their business to
another provider. Reputation is an asset that is expensive to build up and that
is invaluable for any company. Thus, we assume that a principal would not risk
to lose its good reputation for a small gain and will consequently rather adhere
to its policy.
Trust based on explicit punishment means that we do not trust the princi-
pal, but rather the underlying legal framework to ensure the principal's proper
behaviour. Here, we explicitly introduce a similar tradeoff as in trust based
on reputation by imposing disciplinary actions such as fines or imprisonment,
depending on the severity of the offence. The short term gain that might be
achieved through a policy violation is supposed to be negated by appropriate
punishment.
Obviously, there are many problems with this approach, such as the enforce-
ment of laws, which is usuahy expensive, quite slow, and sometimes very complex
(in particular if the laws of different countries are applicable as can be expected
for transactions on the Internet). The difficulty of very different perceptions
of punishment, where a person who has not much to lose might readily risk
some years of imprisonment for the possibility of a relatively large gain. Another
problem in the optimistic approach stems from the fact that many abuses of
confidential information are not necessarily conducted for the purposes of the
company that holds this information, but rather by malicious insiders of such
a company, who do it for strictly personal reasons or financial benefits [20,25].
Such abuses are even more difficult to discover (there are less people involved)
and to punish (it has to be decided if only the employee for malicious behaviour,
only the company for neghgence, or both have to be pursued).
The problem to reliably discover a policy violation could be resolved by
requiring a high degree of transparency. However, this is difficult to achieve and it
is quite likely that even trustworthy principals with a good reputation might not
 476

be eager to accept it in order to protect internal business processes. We therefore

assume that complete transparency is not a very useful tool for supervision.
A better approach would be to designate specialized appraisal companies that
execute frequent in-depth controls of the conduct of companies.
Finally, by definition, the optimistic approach cannot prevent malicious be-
haviour, but it only tries to compensate for it after it has been discovered. For
many situations in real-life, where a violation might have an irreparable effect or
where a proper functioning of the system is absolutely essential, this guarantee
might not be strong enough. A more detailed discussion of the notion of trust
described above, can be found in the paper of Swarup and Fabrega [19].
We would like to remark that most of the described problems are also present
in our every-day life and therefore quite well understood. However, the question
stands if we can do better than that.

3.2 The pessimistic approach

The pessimistic approach removes all these disadvantages by simply preventing

any violation of the published policy. This would clearly be the best foundation
for trust since we can solely rely on a principal's policy to verify that its be-
haviour will be acceptable. The behaviour of the principal becomes completely
transparent as far as it is constrained by its policy without the need to actually
supervise any particular action. If the policy prescribes an action for some event
and if the policy is enforced then it is guaranteed that the action will take place.
Unfortunately, this policy enforcement can not be realized in its full generality,
but is limited to those policies (or rules of a policy) that can effectively be en-
forced with some uncircumventable mechanism. For non-enforceable policies, we
still have to rely on optimistic approaches to trust.
There is no simple way to conceive such an uncircumventable enforcement
mechanism. Until recently it was considered impossible without relying on some
piece of trusted and tamper-resistant hardware [4]. However, in [16] Sander and
Tschudin describe a new approach that might eventually be capable to provide
some protection for an agent without relying on such hardware. Unfortunately,
in its current form the approach does not allow to create agents that encode ar-
bitrary programs, but it only supports polynomial and rational functions. Thus,
the only viable way that can be implemented with current knowledge has to rely
on trusted and tamper-resistant hardware. In the following section, we will de-
scribe such a piece of hardware and the requirements that have to be met so that
it can be used to enforce certain rules of a policy. This hardware is comparable
to the Secure Coprocessor described by Yee in [27].

4 Tamper-resistant hardware and the CryPO protocol

We will first present the execution environment that we rely on and then describe
the protocol that uses it. Figure 1 gives an overview of the principals in the
system.
 477

^^
•^- \_/, ^^^ provides

^° 1 1
' ^^ )oon°pu,er| 1[^^ ^
©
Fig. 1. Overview of the Principals in the CryPO protocol

A manufacturer (TM) produces the execution environment(TPE), which can

be bought by any agent executor (AE). An agent ovi^ner (AO) has to trust
the manufacturer to design and produce its execution environments properly
(see Section 6). The broker (Br) is basically a directory service to locate other
principals and to obtain their credentials.

4.1 Notation
The described approach relies on public key cryptography [5] (such as RSA [14]).
A detailed description of cryptography and the corresponding notations is not
within the scope of this presentation. For information on this topic see e.g., [12,
17]. The notation we will use is as follows.
A principal P has a pair (or several pairs^) of keys {Kp,Kp^) where Kp is
P's public key and Kp^ its private key. Given these keys and the corresponding
algorithm, it is possible to encrypt a message m using the receiver P's public
key Kp, denoted {mJKp, so that only P can decrypt it with its private key. A
signed message, including a digital signature on the message m, generated by P
using its private key Kp^ and verifiable by anybody using the respective public
key Kp, is denoted {m}Sp-
in the following we assume the usage of optimization schemes such as encrypt-
ing a large message with a symmetric session key, which in turn is encrypted
using public key cryptography and prepended to the message as well as the use
of hash algorithms to reduce the computational complexity of signing. However,
for ease of presentation, we will not make this explicit.

4.2 The processing environment

As we have noted above, there is no way to enforce any particular behaviour from
another principal without a piece of trusted and tamper-resistant hardware. The
concept of tamper-resistance usually applies to a well-defined module, sometimes
called black-box, that executes a given task. The outside environment cannot in-
terfere with the task of this module other than through a restricted interface that
' It is advisable to have at least two pairs of keys, one for encryption/decryption and
one for digital signatures.
 478

is under the complete control of the tamper-resistant module. We will call this
device trusted processing environment (TPE). The TPE (see Figure 2) provides
a complete agent platform that cannot be inspected or tampered with. Any
agent residing on the TPE is thus protected by the TPE both from disclosure and
manipulation.

Fig. 2. The trusted processing environment (TPE)

The TPE is a complete computer that consists of a CPU, RAM, ROM, and non-
volatile storage (e.g. hard-disk or flash RAM). It runs a virtual machine (VM)
that provides the platform for the execution of agents and guarantees the correct
execution of the agent's code according to the definition of the used language
(e.g., Java byte-code). Below the VM is the operating system that provides the
external interface to the TPE and controls the VM (e.g., protection of agents from
each other). Furthermore, the TPE contains a private key i^^pg that is known
to no principal other than the TPE - also the physical owner of the TPE does
not know the private key. This can, for instance, be achieved by generating the
private key on the TPE^. Using this approach, the private key is never available
outside of the TPE and, thus, protected by the operating system and the tamper-
resistance of the TPE. The secrecy of the private key is a crucial requirement for
the usage of the TPE to enforce a particular behaviour.
The TPE is connected to a host computer that is under the control of the
TPE's owner. This host computer can access the TPE exclusively through a well
defined interface that allows, for instance, the following operations on the TPE:

- upload, migrate, or remove agents;

- facilitate interactions between host and agent or between agents on the TPE;
- verify certain properties of the TPE (such as which agents are currently
executing).

Due to its implementation as a tamper-resistant module and the restricted

access via the I/O interface, it is impossible to directly access the information
^ Other, more sophisticated approaches to create the pair of keys could be envisaged,
which could also incorporate key recovery mechanisms (e.g., escrowed key shares).
 479

that is contained in the TPE. This property is ensured by the TPE manufacturer
(TM), which also provides the agent executor {AE) with a certificate (signed
by TM). The certificate contains information about the TPE, such as its manu-
facturer, its type, the guarantees provided, and its pubhc key. The agent owner
{AO) has to trust the TM (see Section 6) that the TPE actually does provide
the protection that is claimed in the certificate.

4.3 CryPO protocol

The CryPO (cryptographically protected objects^) protocol transfers agents ex-

clusively in encrypted form over the network to a TPE, using the TPE's public
key. Therefore, it is impossible for anyone who does not know the private key to
obtain the code or data of such a protected agent.
The protocol is divided into two distinct phases. The first phase consists of an
initialization, which has to be executed once before the execution of the second
phase of the protocol. This second phase is concerned with the usage of the TPE
and the actual transfer of the agent. The protocol is based on the interactions
given in Figures 3 and 4.

Initialization In the initiahzation phase, the participants exchange the required

key information:

- we assume that the AO holds an authentic copy of the TM's certification

key KTM-
- the TM sends the certificate CertrpE = {KTPE, •••}STM to the AE.
- the AE registers its reference Ref^^^ with one or several brokers.

Usage After the participants have finished the initialization, they can execute
the usage part of the CryPO protocol:

- the AO queries the broker for the reference to the AE with which it wants
to interact (or it already holds this reference from a previous interaction).
- the AO verifies the policy of the AE whether it is acceptable as well as the
certificate CertrPE to check the manufacturer and the type of the TPE, in
order to decide if it satisfies the security requirements of the AO. If any of
these checks fail, the AO will abort the protocol.
- the AO sends the agent encrypted with the public key of the TPE, {A}KTrEJ
to the AE.
^ We had originally chosen the term object since it is more general than the term
agent.
"* A reference to an AE consists of its name, its physical address in the network, its
policy, and the certificate CertrPE for its TPE. The broker can also verify that the AE
actually controls the corresponding TPE by executing a challenge-response protocol
with the TPE via the AE.
 480

s-sS^^^^IP^^^^^^^

TPE: Trusted Processing Environment AO: Agent Owner

TM: TPE Manufacturer AE: Agent Executor
Ret: Reference for an AE Br: Broker

Fig. 3. Initialization of the CryPO protocol

the AE cannot decrypt {A}KTrB ^^^ can it do anything other than upload
the agent to its TPE.
the TPE decrypts {A}KTrE using its private key Kl^p^ and obtains the exe-
cutable agent A, which it will eventually start. The agent can then interact
with the local environment of the AE or with other agents on the TPE.
the agent can, after it has finished its task, migrate back to its owner
{{A}KAO) or to another AE to which it holds a reference.

^^^^^Srir^^^^^^

TPE: Trusted Processing Environment AO: Agent Owner

TM: TPE Manufacturer AE: Agent Executor
Ref: Reference tor an AE Br: Broker

Fig. 4. Usage of the CryPO protocol

The obvious problem of protecting the TPE from malicious agents is inde-
pendent of the described approach and has to be tackled with additional mecha-
nisms, such as code signing and sandboxing. The problem of protecting the TPE
from tampered agents can easily be solved by concatenating the agent with a
hash of the entire agent h{A), including its execution state, before encrypting
 481

it {A, h{A)}KTPE- The TPE simply has to verify the correct hash before starting
the agent.

4.4 N o t e s on feasibility

The actual construction of a tamper-resistant module in the real world is difficult;

nevertheless, there are many applications that rely on them (e.g., payphones,
debit cards, or SIM cards for GSM). Given sufficient time and resources, it
becomes very probable that an attacker can violate the protection of such a
module (see e.g., [1]). We believe that the actual reahzation of the presented TPE
with reasonably strong guarantees in real-world settings is also quite difficult,
but nonetheless feasible. Especially, since we do not require the prevention but
only the detection of tampering with the TPE for most envisioned applications.
We imagine the TPE as a regular computer with a special operating system.
It is physically protected with a special hardware that can effectively be sealed to
detect tampering, is under continuous video surveillance similar to the systems
used to supervise automatic teller machines, and is subject to challenge inspec-
tions by the TM or an independent appraisal and inspection organization. As
explained in [1], such an installation is conceivable and can even resist massive
attacks. A thorough analysis of the remaining risks has to be undertaken, but
this is not within the scope of this presentation.

5 Usage of t h e TPE

The CryPO protocol together with the concept of a TPE guarantee the integrity
of the agent platform to the AO and protect the code and data of an agent against
manipulation and disclosure, both in transit and during execution. These guar-
antees are based on the trust relation between the AO and the TM, in which the
AO trusts the TM to properly manufacture its TPEs and to control them regu-
larly (if necessary) so that the claimed guarantees hold. The certificate enables
the AO to ensure that it really deals with a TPE from a certain manufacturer.
The above guarantees can be extended by additional properties, formulated
as rules of a policy, that can effectively be enforced by a TPE. In [24], we have
discussed how this approach can be used to allow an agent to base its execution
on results of possible previous executions on the same TPE. This can, for instance,
be used to limit the number of times an agent can be executed on a given TPE. To
achieve this, it is necessary to identify a policy that provides sufficient support
for the agent and to ensure that this policy is enforced on the TPE on which the
agent executes. With this approach, the AO does not need to trust the AE on
the proper protection of his agent, but it suffices to trust the TM. The question
why the AO should trust the TM rather than the AE is discussed in section 6.
Now we want to address a problem that requires the cooperation with a
trusted third party (TTP). Many security related protocols, in particular those
that deal with non-repudiation, rely on such a cooperation [12]. The role of the
TTP is to provide a well defined functionality (e.g., timestamping or logging of
 482

messages) to create non-repudiable evidence that can later be used to resolve a

dispute. The special character of this functionality requires that it is provided
on a trustworthy environment, which can be guaranteed in the administrative
domain of the TTP. Thus, a user of this functionality has to interact with the
TTP server via a remote interaction, which suffers from the usual problems of
limited bandwidth, high latency, and inflexibility of the interface. In order to get
rid of this remote message exchange we propose to encapsulate the functionality
of the TTP in a TTP agent that can be executed on a TPE, which is also a
trustworthy environment. This allows us to gather all interacting parties on a
single platform and to take full advantage of the mobile agent paradigm (see
Section 5.4).
We will first introduce a policy for a TPE and discuss its effect on the agent
of a regular user at the example of a shopping agent. Then we will show how
the same concepts can be applied to a TTP agent and discuss the implications
of the approach.

5.1 The policy of the TPE

We assume that the TPE of the service provider enforces the following set of
rules, detailed in its policy:

a) the code of an agent will never be disclosed or altered by the TPE.

b) any invocation of the agent's methods will be executed exactly according to
the code in the agent.
c) the data of an agent can exclusively be accessed and manipulated through
the interface of the agent. If the agent does not provide methods to directly
access a particular data item, its value can at most be inferred from the
responses to other method invocations.
d) an agent is protected against interference from other agents executing on the
same TPE (other than calls on its public interface).
e) prior to a migration, an agent will obtain the certificate of the designated
receiver's TPE that also contains the policy. The agent can decide whether it
wants to be transferred and the current TPE will honour the agent's decision.
The actual transfer follows the CryPO protocol.
f) the TPE provides an internal clock with reasonable accuracy (on the order
of several seconds). It will try to synchronize this clock with a trusted time
service and will inform the agent if this synchronization did not succeed.

The first rules a) and b) guarantee the basic protection of the agent's code
as well as its proper execution, while c) guarantees the protection of the agent's
data from undesired disclosure and manipulation. Rule d) requires the protection
of agents from one another, which is a regular operating system functionality.
The next rule e) ensures that the agent knows the policy of the TPE to which
it is transferred. Thus, the agent can ensure that it will not be sent to a TPE
that provides insufficient protection. Finally, rule f) can be used for several
purposes. For instance, it allows an agent that contains an expiration date to
 483

implement a limited lifetime (on the order of a few days or hours). Upon its
arrival the agent requests the current time and checks if this time is still within
its attributed lifetime. If its expiration date has passed or if the TPE did not
succeed to synchronize its clock, the agent can simply abort. An AE can not
prevent this if the code of the agent is protected and if it will be executed
correctly.

5.2 The shopping agent

Consider a shopping agent that searches for a particular service for its owner.
Once it has found a suitable offer, it will negotiate the details of the service
provision, such as exact price and various QoS parameters, with the service
provider. As a special requirement, we specify that the shopping agent has to
create a log entry with a TTP server^ that contains the details of the negotiated
contract before providing the payment data. This allows the AO to reconstruct
the activities of his agent in the case of a dispute or if the agent is lost.
In order for the shopping agent to effectively conduct a negotiation it needs
to conceal some of its configuration information from the service provider, such
as the highest acceptable price or the lowest acceptable QoS parameters. Fur-
thermore, the shopping agent holds the public key of the TTP, which it needs
to verify the acknowledgement from the TTP server, as well as the payment
data, which should only be provided to the selected service provider after the
successful creation of a log entry.
If the shopping agent executes only on TPEs that enforce the policy discussed
in Section 5.1, it is clear that it is protected from any interference. Provided
that the agent is correct, no other entity will be able to access or manipulate
any data contained in the agent other than what is accessible via the methods
of its public interface. Thus, the agent can effectively negotiate with a service
provider, request the logging of the contract with a TTP server, and delay any
further actions until it has received a signed acknowledgement from the TTP
server.

5.3 The T T P agent

The interaction described above allows for an efficient negotiation between the
agent and the service provider exploiting all the performance advantages of the
mobile agent paradigm. However, due to the special requirements of the AO, the
agent has to interact with a TTP server via a remote interaction. Apart from
the performance penalties of this remote interaction, the TTP server can also
become a bottleneck if its resources are consumed by a large number of clients.
Therefore, we propose to encapsulate the functionality of the TTP in a TTP

^ The agent could send the corresponding information directly to the AO, but since
it needs an acknowledgement for the receipt of the log message and since the AO
might not have a permanent connection to the network, it is preferable to delegate
this task to a TTP.
 484

agent (TA) that can be executed on the TPE, relying on the same protection
mechanisms as the shopping agent®.
In the case of message logging, the functionality of the TTP consists of ac-
cepting arbitrary messages, storing them up to a well defined point in time t,
and responding with an acknowledgement asserting that the message has been
logged. This acknowledgement has to be signed by the TTP and must clearly
identify the message that was supposed to be logged, either by including the
message itself or, preferably, a hash of the message. Furthermore, the TTP must
be capable to reproduce a logged message up to the time t and to provide it
(exclusively to authorized principals) upon request. If necessary, a log message
can be confidentiality protected with regular encryption methods.
The task of the TA is to act as the proxy for the actual TTP on the TPE. It
will accept messages that have to be logged from agents on the TPE, store them
in a local cache, and respond with an acknowledgement in which it guarantees
that it will forward the message to the TTP server unless the TPE is destroyed
(see below). Once a log message arrives at the TTP server, it will be handled
like a normal message. In order for the TA to provide such a guarantee, it needs
access to a sufficient amount of non-volatile storage on the TPE, in which it can
safely store the log messages. Since this non-volatile storage is a limited resource
of the TPE, the TA needs a special authorization to use it. If the TPE owner does
not grant this authorization, the TA will abort.
Apart from the access to the non-volatile storage of the TPE, the requirements
of the TA are very similar to those of the shopping agent. It has to be protected
against manipulation of its code and data and it also has to conceal certain data
items from the agent executor such as two different cryptographic keys. The first
key is necessary as a means to securely forward the logged messages to the TTP
server. Since the TA is configured by the TTP, this can simply be a secret key of
a symmetric key cryptosystem. The second key is needed as a signature key to
sign the acknowledgements for logged messages. This does not necessarily have
to be the long-term signature key of the TTP, but can be a temporary key that
is validated by a certificate signed with the TTP's long-term signature key.
Again, if the T T P ensures that the TA only executes on TPEs that enforce
the policy discussed in Section 5.1 with sufficiently high assurance, it is clear that
the keys are protected and that the proper execution of the TA is guaranteed.
The remaining problem is how the TA can guarantee that logged messages that
are stored in the non-volatile storage of the TPE will eventually be forwarded to
the TTP server. The TPE owner could simply intercept all the messages of the
TA to the TTP server or, ultimately, request the TPE to terminate the TA. This
functionality has to be offered by the TPE to protect its owner from malicious
or simply buggy agents that refuse to terminate.
This problem can be solved with a supervision of the TA by the TTP and
with the help of the internal clock provided by the TPE. The TTP has to keep

A TTP might require a higher level of assurance in the protection of the TPE than
a regular user. This could be a differentiating feature of TPEs from different manu-
facturers.
 485

track of all the TAs it sent to the various service providers and of an expiration
date that is associated with each TA. A TA will accept log messages only until
its expiration date. After this date it will refuse to accept and acknowledge any
further messages. Thus, the TTP has to receive a final message from the TA
after its expiration date (there should be an additional delay to accommodate
for clock skew) indicating that no further messages for the TTP are stored in
the non-volatile storage of the TPE. Provided that the TA only deletes mes-
sages from this non-volatile storage after sending them to the TTP server and
obtaining an acknowledgement, the TTP knows that all the messages that the
TA acknowledged have been forwarded to the TTP server. If this final message
is not received, the TTP will request the TPE owner to restart the TA and to
forward any messages it sends to the TTP server. Under the assumption that
the TTP has a possibility to enforce the access to the TPE by legal means, the
only possibility of the TPE owner to avoid the provision of missing messages is to
destroy the TPE. Thus, the approach cannot guarantee that all logged messages
will be delivered to the TTP server. But it can guarantee that a TPE owner can
not cheat without being discovered. Furthermore, if it can be proven that the
TPE owner intentionally destroyed the TPE, he can be punished with adequate
fines.
The problem of destruction of storage media is not new and also applies
to a regular TTP. However, it is assumed that the TTP operator implements
adequate measures to avoid this problem.

5.4 Discussion
The relocation of the TTP functionality from a remote site to the locally man-
aged TPE allows us to prevent it from becoming a bottleneck that slows down
other components. This is possible since the TPE owner can allocate as many re-
sources as necessary to the TA without having to coordinate this with the TTP.
Also, since the TA can collect and merge several log messages from interactions
of different agents with the service provider, it can accumulate several log mes-
sages and forward them in a single remote interaction. Another major advantage
of the described approach is that the remote interaction with the TTP server
is taken off the critical communication path between the agent and the service
provider. They can continue their interaction as soon as the TA has stored the
log message and sent the acknowledgement. Moreover, the entire interaction can
exploit the locally available communication links with higher bandwidth and
lower latency. This enables not only a better overall performance of the system,
but allows interactions that were not possible before due to an unreasonable
overhead. For instance, a TA could be used by two interacting parties as an
intermediate through which all messages are exchanged. The TA could, thus,
easily log the entire interaction and forward it to the TTP.
The concept of the TA is suitable for any TTP functionality that can be
wrapped up in a reasonably small object and that does not have to rely on a
large centrally managed database. Other interesting examples are timestamp-
ing or fair-exchange. The former consists of a TA that adds a timestamp to a
 486

message and signs the resulting message with its signature itey. The latter is a
classical security problem, for which several solutions relying on a TTP have
been proposed [2]. The problem of fair-exchange is that of principals A and B
who want to exchange the data items DA and DB , but neither of them wants
to provide its data item before receiving that of the other principal. A TA can
facilitate the exchange by accepting the data items as well as a description of the
data items expected by the designated receivers. It will verify if the descriptions
match the actual data items (e.g., in the case of payment, it verifies if the paid
amount corresponds with the amount expected by the receiver) and, if this is
the case, deliver the data items to the designated receiver.

6 Trust in t h e TPE manufacturer

We have introduced the mechanism, with which an agent can take advantage of
the policy enforced by a TPE. However, as we have mentioned above, in order
for a principal to trust in the proper enforcement of this policy, it is necessary
that he also trusts the TPE manufacturer to properly design, implement, and
produce its TPEs. Since there is no way (to the knowledge of the authors) to
enforce a correct behaviour of the TPE manufacturer, it seems that the presented
approach simply replaces one required trust relationship with another one. This
is a correct observation from a theoretical point of view. Nevertheless, we believe
that the replacement of trust in an arbitrary service provider with trust in a TPE
manufacturer has several more subtle implications. We will briefly discuss the
following advantages that we identified:

- better understanding of security and privacy problems

— centralized control
- resources to build reputation
— separation of concern

The TPE manufacturer is a specialized service provider, which primarily deals

in the field of the provision of security devices. Therefore it has a better under-
standing of security and privacy problems, which makes it a much more capable
entity to ensure this service since it is more aware of the potential problems and
pitfalls.
We assume that there will be relatively few TPE manufacturers (on the order
of several hundreds) compared to the number of possible operators of the TPE
(on the order of several millions). This makes the control of their behaviour much
easier for expert appraisal organizations. Also, it is quite conceivable that a TPE
manufacturer might invite external experts to control its internal operation, in
order to obtain a better position in the market (similar to the approach for
quality assurance in the ISO-9000).
The production of TPEs is considered to be a difficult task (see Section 4.4).
Therefore, we assume that it will be undertaken by major corporations, which
have the necessary resources to build a good reputation and which have an
 487

incentive to protect this reputation. This allows us to rely on good reputation

as foundation for trust in the TPE manufacturer.
The TPE manufacturer that is responsible for the enforcement of the proper
pohcy rules on the TPE, has nothing to gain by not accomphshing its task.
Since the TPE will be operated independent from the TPE manufacturer by a
completely different principal and since the TPE manufacturer has no means to
access the data that is processed on the TPE (no physical connection), there is
no possibility for the TPE manufacturer to draw a direct benefit from a TPE that
does not properly enforce its policy^.
We assume that the above arguments of high expertise, effective controlla-
bility, good reputation, and lack of incentive are sound reasons to trust a TPE
manufacturer to build reliable and powerful TPEs. The main advantage of the
approach lies in the possibility to leverage this trust in the TPE manufacturer
onto a completely different principal in the role of the service provider, which

— might not have the proper expertise to ensure a secure operation of its hard-
ware and to guarantee the protection of the processed data.
— is quite difficult to control, due to the sheer number of such service providers.
— might have no particular reputation (and therefore none to lose).
— might have short term goals that (in its point of view) justify a policy vio-
lation.

With the presented approach, such a service provider can easily define the
pohcy rules that it would like its TPE to enforce (by selecting from the options
offered by the TPE manufacturer) and buy the appropriate TPE from a reputable
TPE manufacturer. The service provider can then immediately benefit from the
trust that users have in the manufacturer of its TPE to convince them that it
will not maliciously abuse an agent sent by a user.
With this, the approach favours the open systems philosophy, where any
principal can possibly become a provider of services. Such a service provider
simply has to obtain a TPE from some reputable manufacturer and can then
easily convince a client that the client's confidential information is sufficiently
well protected. Thus, it becomes much easier for a new service provider to es-
tablish itself in the market.

7 Conclusion

In this paper, we have discussed the notion of trust in the context of mobile agent
systems and introduced a structuring for this problem domain. Starting from this
structure, we have proposed an approach that relies on a trusted and tamper-
resistant hardware device, which allows the prevention of malicious behaviour
rather than its correction. We believe this to be the better form of protection for

'' There is the possibihty that a TPE operator bribes a TPE manufacturer to provide
an incorrect TPE. We assume that such a behaviour is a severe offence that is subject
to criminal investigation and not within the scope of this discussion.
 488

confidential data. We have shown how the approach can be used to effectively
protect the confidential data contained in the shopping agent of a user and
how it can be extended to protect specialized agents from TTPs that provide
facilitation services.
In real-life, there are limitations to the approach. Given sufficient time and
resources, a TPE operator might succeed in breaking the system and it would
thus be possible for him to violate the policy that should be enforced by the TPE.
Our goal is to make this attack so costly that it would negate a possible gain
(there may be many different implementations that provide different levels of
assurance in the protection of a TPE). As further deterrent, we assume that a non-
repudiable proof for a policy violation of an enforced policy or for an attempted
or successful breaking of a TPE might be punished much more severely than a
mere policy violation since it proves a much larger determination to commit a
criminal offence.

Acknowledgements

This research was supported by a grant from the EPFL ("Privacy" project) and
by the Swiss National Science Foundation as part of the Swiss Priority Pro-
gramme Information and Communications Structures (SPP-ICS) under project
number 5003-045364.

References
1. R. Anderson and M. Kuhn. Tamper resistance — a cautionary note. In The Second
USENIX Workshop on Electronic Commerce Proceedings, pages 1-11, Oakland,
California, November 1996.
2. H. Biirk and A. Pfitzmajin. Value exchange systems enabling security and unob-
servability. Computers & Security, 9(8):715-721, 1990.
3. A. Carzaniga, G. P. Picco, and G. Vigna. Designing distributed applications with
mobile code paxadigms. In R.Taylor, editor. Proceedings of the 19th International
Conference on Software Engineering (ICSE'97), pages 22-32. ACM Press, 1997.
4. D. M. Chess, B. Grosof, C. G. Harrison, D. Levine, C. Parris, and G. Tsudik.
Itinerant agents for mobile computing. IEEE Personal Communications, 2(3):34-
49, October 1995.
5. W. Diffie and M. E. Hellman. New directions in cryptography. IEEE Transactions
on Information Theory, IT-22(6), November 1976.
6. DoD. Trusted Computer System Evaluation Criteria (TCSEC). Technical Report
DoD 5200.28-STD, Department of Defense, December 1985.
7. J. Gosling and H. McGilton. The Java language environment. White paper. Sun
Microsystems, Inc., 1996.
8. R.S. Gray. Agent Tel; A transportable agent system. In Proceedings of the CIKM
Workshop on Intelligent Information Agents, Baltimore, MD, December 1995.
9. C. G. Harrison, D. M. Chess, and A. Kershenbaum. Mobile agents: Are they a
good idea? In Mobile Object Systems: Towards the Programmable Internet, volume
1222 of Lecture Notes in Computer Science, pages 25-47. Springer Verlag, 1997.
 489

10. ITU. ITU-T Recommendation X.509: The Directory - Authentication Framework.

International Telecommunication Union, 1993.
11. D. B. Lange and M. Ishima. Program and Deploying Java Mobile Agents with
Aglets. Addison-Wesley, 1998.
12. A. J. Menezes, P. C. van Oorschot, and S. A. Vanstone. Handbook of applied
cryptography. CRC Press, Inc., 1997.
13. J. Ordille. When agents roam, who can you trust? Technical Report Technical
Report, Computing Science Research Center, Bell Labs, 1996.
14. RSA Data Security, Inc. PKCS #1: RSA Encryption Standard. RSA Data Security,
Inc., November 1993.
15. R. A. Rueppel. A formal approach to security architectures. In EuroCrypt, pages
387-398, Brighton, England, 1991.
16. T. Sander and C. Tschudin. Towards mobile cryptography. In IEEE Symposium
on Security and Privacy, May 1998.
17. B. Schneier. Applied cryptography. Wiley, New York, 1994.
18. J. G. Steiner, C. Neuman, and J. I. Schiller. Kerberos: An authentication service
for open network systems. In Proceedings of the USENIX Winter 1988 Technical
Conference, pages 191-202. USENIX Association, Berkeley, USA, February 1988.
19. V. Swarup and J. T. Fabrega. Understanding trust. In Secure Internet Program-
ming [22].
20. New York Times. U.S. workers stole data on 11,000, agency says, April 6, 1996.
21. G. Vigna. Protecting mobile agents through tracing. In Proceedings of the Third
Workshop on Mobile Object Systems, Finland, June 1997.
22. Jan Vitek and Christian Jensen. Secure Internet Programming: Security Issues
for Mobile and Distributed Objects. Lecture Notes in Computer Science. Springer-
Verlag Inc., New York, NY, USA, 1999.
23. J. E. White. Telescript technology: The foundation for the electronic market place.
White paper. General Magic, Inc., 1994.
24. U. G. Wilhelm, L. Buttyan, and S. Staamann. On the problem of trust in mobile
agent systems. In Symposium on Network and Distributed System Security, pages
114-124. Internet Society, March 1998.
25. I. S. Winkler. The non-technical threat to computing systems. Computing Systems,
USENIX Association, 9(1):3-14, Winter 1996.
26. T. Y. C. Woo and S. S. Lam. Authentication for distributed systems. IEEE
Computer, 25(l):39-52, January 1992.
27. B. Yee. A sancturary for mobile agents. In Secure Internet Programming [22].
28. P. Zimmermann. PGP User's Guide. MIT Press, Cambridge, 1994.
 Part IV

Appendix
 List of Authors (Feb 1999)

Foundations

Martin Abadi
Systems Research Center
Compaq
130 Lytton Avenue
Palo Alto, CA 94301
U.S.A.

Massimo Ancona
DISI - University of Genova
Via Dodecanese 35, 16146 Genova
Italy
[email protected]
http://www.disi.unige.it/person/AnconaM/

Luca Cardelli
Microsoft Research
1 Guildhall St, Cambridge CB2-3NH
United Kingdom
lucaSluca.demon.co.uk
http://www.luca.demon.co.uk

Walter Cazzola (DSI - University of Milano)

C/O DISI - University of Genova,
Via Dodecaneso 35, 16146 Genova
Italy
cazzolaSdsi.unimi.it
http://www.disi.unige.it/person/CazzolaW/

Rocco De Nicola
Dipartimento de Sistemi e Informatica
Universita di Firenze
Via Lombroso 6/17,1-50134 Firenze
Italy
denicolaSdsi.unifi.it
http://dsi2.dsi.unifi.it/~denicola
 494

Eduardo B. Fernandez
Department of Computer Science and Engineering
Florida Atlantic University
777 Glades Road, Boca Raton, Florida 33431 U.S.A.
edOcse.fau.edu
http://www.cse.fau.edu/"ed/

GianLuigi Ferrari
Dipartimento di Informatica
Universita di Pisa
Corso Italia 40,1-56125 Pisa
Italy
[email protected]
http://www.di.unipi.it/"giangi

Matthew Hennessy
School of Cognitive and Computer Science
University of Sussex
Falmer, Brighton BNl 9QH
United Kingdom
[email protected]
http://www.cogs.susx.ac.uk/users/matthewh/

Xavier Leroy
INRIA Rocquencourt
Domaine de Voluceau, 78153 Le Chesnay
France
Xavier.LeroyOinria.fr

Rosario Pugliese
Dipartimento de Sistemi e Informatica
Universita di Firenze
Via Lombroso 6/17,1-50134 Firenze
Italy
puglieseSdsi.unifi.it
http://dsi2.dsi.unifi.it/"pugliese

James Riely
Department of Computer Science
North Carolina State University
Raleigh, NC 27695-7534
U.S.A.
rielyScsc.ncsu.edu
http://www.CSC.ncsu.edu/faculty/riely
 495

Francois Rouaix
Liquid Market Inc
5757 West Century Bid
Los Angeles CA 90045
U.S.A.

Vipin Swarup
The MITRE Corporation
202 Burlington Road
Bedford, MA 01730
U.S.A.

Javier Thayer
The MITRE Corporation
202 Burlington Road
Bedford, MA 01730
U.S.A.

Concepts

Tuomas Aura
Helsinki University of Technology
Laboratory for Theoretical Computer Science
P.O. Box 1100, FIN-02015 HUT
Finland
[email protected]
http://www.tcs.hut.fi/

Matt Blaze
AT&T Labs - research
180 Park Avenue
Florham Park, NJ 07932
U.S.A.
[email protected]

Gerald Brose
Institut fiir Informatik
Freie Universitat Berlin
TakustraBe 9, D-14195 Berlin
Germany
[email protected]
http://www.inf.fu-berlin.de/"brose
 496

Joan Feigenbaum
AT&T Labs - research
180 Park Avenue
Florham Park, NJ 07932
U.S.A.
j f©research.att.com

John loannidis
AT&T Labs - research
180 Park Avenue
Florham Park, NJ 07932
U.S.A.
j [email protected]

Angelos D. Keromytis
Distributed Systems Lab
CIS Department University of Pennsylvania
200 S. 33rd Str.
Philadelphia, PA 19104
U.S.A.
angelosSdsl.cis.upenn.edu

Christian F. Tschudin
Department of Computer Systems
Uppsala University
Box 325, SE - 751 05 Uppsala
Sweden
tschudinSdocs.uu.se
http://www.docs.uu.se/ tschudin/

Bennet S. Yee
Department Computer Science and Engineering, 0114
University of California San Diego
9500 Oilman Dr
La Jolla,CA 92093-0114
U.S.A.
[email protected]
http://www.cse.ucsd.edu/~bsy/

Volker Roth
Fraunhofer Institut fiir Graphische Datenverarbeitung
RundeturmstraBe 6, D-64283
Germany
[email protected]
http: //www. igd. fhg.de/"'vroth
 497

Implementations

Martin Abadi
See contact information above.

Anurag Acharya
Department of Computer Science
University of California
Santa Barbara, CA 93106
U.S.A.
[email protected]

D. Scott Alexander
Bell Labs, Lucent Technologies
600 Mountain Avenue
Murray Hill, NJ 07974
U.S.A.
salexSresearch.bell-labs.com

William A. Arbaugh
Distributed Systems Lab
CIS Department University of Pennsylvania
200 S. 33rd Str.
Philadelphia, PA 19104
U.S.A.
waaOdsl.cis.upenn.edu

Brian N. Bershad
Department of Computer Science and Engineering
University of Washington
Box 352350
Seattle, WA 98195
U.S.A.
[email protected]
http://www.cs.washington.edu/homes/bershad

Mike Burrows
Systems Research Center
Compaq
130 Lytton Avenue
Palo Alto, CA 94301
U.S.A.
 498

Levente Buttyan
Institute for computer Communications and Applications
Swiss Federal Institute of Technology (EPFL)
1015 Lausanne
Switzerland
[email protected]
http://icawww.epf1.ch/buttyan

Chi-Chao Chang
Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
http://simon.cs.Cornell.edu/Info/People/chichao/chichao.html

Vipin Chaudhary
Depeartment of ECE
Wayne State University
Detroit, MI 48202
U.S.A.
vipinOece.eng.wayne.edu

Grzegorz Czajkowski
Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
http://simon.cs.Cornell.edu/home/grzes/

Leendert van Doom

IBM Thomas J. Watson Research Center
30 Saw Mill River Rd.
Hawthorne, NY 10532
U.S.A.
[email protected]

Guy Edjlali
Depeartment of ECE
Wayne State University
Detroit, MI 48202
U.S.A.
[email protected]
 499

Thorsten von Eicken

Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
http://www.cs.Cornell.edu/tve/

Robert Grimm
Department of Computer Science and Engineering
University of Washington
Box 98195
Seattle, WA 98195
U.S.A.
[email protected]
http://www.cs.Washington.edu/homes/rgrimm

Jiirgen Harms
Centre Universitaire d'Informatique
University of Geneva
Rue General Dufour 24
1211 Geneve 4
Switzerland
Juergen.HarmsOcui.unige.ch

Chris Hawblitzel
Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
http://simon.cs.Cornell.edu/Info/People/hawblitz/hawblitz.html

Deyu Hu
Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
http://simon.cs.Cornell.edu/Info/People/hu/hu.html
 500

Jarle G. Hulaas
Centre Universitaire d'Informatique
University of Geneva
Rue General Dufour 24
1211 Geneve 4
Switzerland
[email protected]

Trent Jaeger
IBM Thomas J. Watson Research Center
30 Saw Mill River Rd.
Hawthorne, NY 10532
U.S.A.
jaegertSwatson.ibm.com

Michael B. Jones
Microsoft Research, Microsoft Corporation
One Microsoft Way, Building 31/2260
Redmond, WA 98052
U.S.A.
mbj ©microsoft.com
http://www.research.microsoft.com/~mb:

Angelos D. Keromytis
See contact information above.

Jonathan M. Smith
Distributed Systems Lab
CIS Department University of Pennsylvania
200 S. 33rd Str.
Philadelphia, PA 19104
U.S.A.
j ms (ids 1. c i s . u p e n n . edu

Dan Spoonhower
Department of Computer Science
Cornell University
Ithaca, NY 14850
U.S.A.
[email protected]
 501

Sebastian Staamann
Operating Systems Laboratory (LSE)
Swiss Federal Institute of Technology (EPFL)
1015 Lausanne
Switzerland
[email protected]
http://Isewww.epf1.ch/~staa
Alex Villazon
Centre Universitaire d'Informatique
University of Geneva
Rue General Dufour 24
1211 Geneve 4
Switzerland
[email protected]

Uwe G. WUhelm
Operating Systems Laboratory (LSE)
Swiss Federal Institute of Technology (EPFL)
1015 Lausanne
Switzerland
Uwe .Wilhelmliepf 1. ch
http://Isewww.epf1.ch/~wilhelm

Edward Wobber
Systems Research Center
Compaq
130 Lytton Avenue
Palo Alto, CA 94301
U.S.A.
Lecture Notes in Computer Science
For information about Vols. 1-1537
please contact your bookseller or Springer-Verlag

Vol. 1538; J. Hsiang, A. Ohori (Eds.), Advances in Com- Vol. 1556: S. Tavares, H. Meijer (Eds.), Selected Areas
puting Science - ASIAN'98. Proceedings, 1998. X, 305 in Cryptography. Proceedings, 1998. IX, 377 pages. 1999.
pages. 1998. Vol. 1557: P. Zinterhof, M. Vajtersic, A. Uhl (Eds.), Par-
Vol. 1539: O. RUthing, Interacting Code Motion Trans- allel Computation. Proceedings, 1999. XV, 604 pages.
formations; Their Impact and Their Complexity. XXI,225 1999.
pages. 1998. Vol. 1558: H. J.v.d. Herik, H. lida (Eds.), Computers and
Vol. 1540; C. Beeri, P. Buneman (Eds.), Database Theory Games. Proceedings, 1998. XVIII, 337 pages. 1999.
- ICDT'99. Proceedings, 1999. XI, 489 pages. 1999. Vol. 1559: P. Flener (Ed.), Logic-Based Program Syn-
Vol. 1541; B. Kigstrom, J. Dongarra, E. Elmroth, J. thesis and Transformation. Proceedings, 1998. X, 331
Wasniewski (Eds.), Applied Parallel Computing. Proceed- pages. 1999.
ings, 1998. XIV, 586 pages. 1998. Vol. 1560: K. Imai, Y. Zheng (Eds.), Public Key Cryp-
Vol. 1542; H.I. Christensen (Ed.), Computer Vision Sys- tography. Proceedings, 1999. IX, 327 pages. 1999.
tems. Proceedings, 1999. XI, 554 pages. 1999. Vol. 1561: I. Damgjrd (Ed.), Lectures on Data
Vol. 1543; S. Demeyer, J. Bosch (Eds.), Object-Oriented Security.VII, 250 pages. 1999.
Technology ECOOP'98 Workshop Reader. 1998. XXII, Vol. 1562; C.L. Nehaniv (Ed.), Computation for Meta-
573 pages. 1998. phors, Analogy, and Agents. X, 389 pages. 1999.
Vol. 1544: C. Zhang, D. Lukose (Eds.), Multi-Agent Sys- (Subseries LNAI).
tems. Proceedings, 1998. VII, 195 pages. 1998. (Subseries Vol. 1563: Ch. Meinel, S. Tison (Eds.), STAGS 99. Pro-
LNAI). ceedings, 1999. XIV, 582 pages. 1999.
Vol. 1545: A. Birk, J. Demiris (Eds.), Learning Robots. Vol. 1565; P. P. Chen, J. Akoka, H. Kangassalo, B.
Proceedings, 1996. IX, 188 pages. 1998. (Subseries Thalheim (Eds.), Conceptual Modeling. XXIV, 303 pages.
LNAI). 1999.
Vol. 1546: B. Moller, J.V. Tucker (Eds.), Prospects for Vol. 1567; P. Antsaklis, W. Kohn, M. Lemmon, A.
Hardware Foundations. Survey Chapters, 1998. X, 468 Nerode, S. Sastry (Eds.), Hybrid Systems V. X, 445 pages.
pages. 1998. 1999.
Vol. 1547; S.H. Whitesides (Ed.), Graph Drawing. Pro- Vol. 1568: G. Bertrand, M. Couprie, L. Perroton (Eds.),
ceedings 1998. Xn, 468 pages. 1998. Discrete Geometry for Computer Imagery. Proceedings,
Vol. 1548; A.M. Haeberer (Ed.), Algebraic Methodology 1999. XI, 459 pages. 1999.
and Software Technology. Proceedings, 1999. XI, 531 Vol. 1569; F.W. Vaandrager, J.H. van Schuppen (Eds.),
pages. 1999. Hybrid Systems; Computation and Control. Proceedings,
Vol. 1549; M. Pettersson, Compiling Natural Semantics. 1999. X, 271 pages. 1999.
XVII, 240 pages. 1999. Vol. 1570: F. Puppe (Ed.), XPS-99: Knowledge-Based
Vol. 1550; B. Christiansen, B. Crispo, W.S. Harbison, Systems. VIII, 227 pages. 1999. (Subseries LNAI).
M. Roe (Eds.), Security Protocols. Proceedings, 1998. Vol. 1571; P. Noriega, C. Sierra (Eds.), Agent Mediated
VIII, 241 pages. 1999. Electronic Commerce. Proceedings, 1998. IX, 207 pages.
Vol. 1551; G. Gupta (Ed.), Practical Aspects of Declara- 1999. (Subseries LNAI).
tive Languages. Proceedings, 1999. VIII, 367 pgages. Vol. 1572; P. Fischer, H.U. Simon (Eds.), Computational
1999. Learning Theory. Proceedings, 1999. X, 301 pages. 1999.
Vol. 1552; Y. Kambayashi, D.L. Lee, E.-P. Lim, M.K. (Subseries LNAI).
Mohania, Y. Masunaga (Eds.), Advances in Database Vol. 1574; N. Zhong, L. Zhou (Eds.), Methodologies for
Technologies. Proceedings, 1998. XIX, 592 pages. 1999. Knowledge Discovery and Data Mining. Proceedings,
Vol. 1553; S.F. Andler, J. Hansson (Eds.), Active, Real- 1999. XV, 533 pages. 1999. (Subseries LNAI).
Time, and Temporal Database Systems. Proceedings, Vol. 1575; S. Jahnichen (Ed.), Compiler Construction.
1997. VIII, 245 pages. 1998. Proceedings, 1999. X, 301 pages. 1999.
Vol. 1554: S. Nishio, F. Kishino (Eds.), Advanced Multi- Vol. 1576; S.D. Swierstra (Ed.), Programming Languages
media Content Processing. Proceedings, 1998. XIV, 454 and Systems. Proceedings, 1999. X, 307 pages. 1999.
pages. 1999.
Vol. 1577; J.-P. Finance (Ed.), Fundamental Approaches
Vol. 1555; J.P. Muller, M.P. Singh, A.S. Rao (Eds.), In- to Software Engineering. Proceedings, 1999. X, 245 pages.
telligent Agents V. Proceedings, 1998. XXIV, 455 pages. 1999.
1999. (Subseries LNAI).
Vol. 1578: W. Thomas (Ed.), Foundations of Software Vol. 1603: J. Vitek, C D . Jensen (Eds.), Secure Internel
Science and Computation Structures. Proceedings, 1999. Programming. X, 501 pages. 1999.
X, 323 pages. 1999. Vol. 1605: J. Billington, M. Diaz, G. Rozenberg (Eds.),
Vol. 1579: W.R. Cleaveland (Ed.), Tools and Algorithms Application of Petri Nets to Communication Networks.
for the Construction and Analysis of Systems. Proceed- IX, 303 pages. 1999.
ings, 1999. XI, 445 pages. 1999. Vol. 1606: J. Mira, J.V. Sanchez-Andres (Eds.), Founda-
Vol. 1580: A. Vdkovski, K.E. Brassel, H.-J. Schek (Eds.), tions and Tools for Neural Modeling. Proceedings, Vol.
Interoperating Geographic Information Systems. Proceed- I, 1999. XXIII, 865 pages. 1999.
ings, 1999. XI, 329 pages. 1999. Vol. 1607: J. Mira, J.V. SSnchez-Andres (Eds.), Engineer-
Vol. 1581: J.-Y. Girard (Ed.), Typed Lambda Calculi and ing Applications of Bio-Inspired Artificial Neural Net-
AppUcations. Proceedings, 1999. VIII, 397 pages. 1999. works. Proceedings, Vol. II, 1999. XXIII, 907 pages. 1999.
Vol. 1582: A. Lecomte, F. Lamarche, G. Perrier (Eds.), Vol. 1609: Z. W. Rui, A. Skowron (Eds.), Foundations of
Logical Aspects of Computational Linguistics. Proceed- Intelligent Systems. Proceedings, 1999. XII, 676 pages.
ings, 1997. XI, 251 pages. 1999. (Subseries LNAI). 1999. (Subseries LNAI).
Vol. 1583: D. Scharstein, View Synthesis Using Stereo Vol. 1610: G. Cornuejols, R.E. Burkard, G.J. Woeginger
Vision. XV, 163 pages. 1999. (Eds.), Integer Programming and Combinatorial Optimi-
Vol. 1584: G. Gottlob, E. Grandjean, K. Seyr (Eds.), Com- zation. Proceedings, 1999. IX, 453 pages. 1999.
puter Science Logic. Proceedings, 1998. X, 431 pages. Vol. 1611: I. Imam, Y. Kodratoff, A. El-Dessouki, M.
1999. All (Eds.), Multiple Approaches to Intelligent Systems.
Vol. 1585: B. McKay, X. Yao, C.S. Newton, J.-H. Kim, Proceedings, 1999. XIX, 899 pages. 1999. (Subseries
T. Furuhashi (Eds.), Simulated Evolution and Learning. LNAI).
Proceedings, 1998. XIII, 472 pages. 1999. (Subseries Vol. 1612: R. Bergmann, S. Breen, M. Goker, M. Manago,
LNAI). S. Wess, Developing Industrial Case-Based Reasoning
Vol. 1586: J. Rolim et al. (Eds.), Parallel and Distributed Applications. XX, 188 pages. 1999. (Subseries LNAI).
Processing. Proceedings, 1999. XVII, 1443 pages. 1999. Vol. 1614: D.P. Huijsmans, A.W.M. Smeulders (Eds.),
Visual Information and Information Systems. Proceed-
Vol. 1587: J. Pieprzyk, R. Safavi-Naini, J. Seberry (Eds.),
ings, 1999. XVn, 827 pages. 1999.
Information Security and Privacy. Proceedings, 1999. XI,
327 pages. 1999. Vol. 1615: C. Polychronopoulos, K. Joe, A. Fukuda, S.
Vol. 1590: P. Atzeni, A. Mendelzon, G. Mecca (Eds.), Tomita (Eds.), High Performance Computing. Proceed-
The World Wide Web and Databases. Proceedings, 1998. ings, 1999. XIV, 408 pages. 1999.
VIII, 213 pages. 1999. Vol. 1617: N.V. Murray (Ed.), Automated Reasoning with
Vol. 1592: J. Stern (Ed.), Advances in Cryptology - Analytic Tableaux and Related Methods. Proceedings,
EUROCRYPT '99. Proceedings, 1999. XII, 475 pages. 1999. X, 325 pages. 1999. (Subseries LNAI).
1999. Vol. 1620: W. Horn, Y. Shahar, G. Lindberg, S.
Vol. 1593: P. Sloot, M. Bubak, A. Hoekstra, B. Hertrberger Andreassen, J. Wyatt (Eds.), Artificial Intelligence in
(Eds.), High-Performance Computing and Networking. Medicine. Proceedings, 1999. XIII, 454 pages. 1999.
Proceedings, 1999. XXIII, 1318 pages. 1999. (Subseries LNAI).

Vol. 1594: P. Ciancarini, A.L. Wolf (Eds.), Coordination Vol. 1621: D. Fensel, R. Studer (Eds.), Knowledge Ac-
Languages and Models. Proceedings, 1999. IX, 420 pages. quisition Modeling and Management. Proceedings, 1999.
1999. XI, 404 pages. 1999. (Subseries LNAI).

Vol. 1596: R. Poll, H.-M. Voigt, S. Cagnoni, D. Corne, Vol. 1622: M. Gonzalez Harbour, J. A. de la Puente (Eds.),
G.D. Smith, T.C. Fogarty (Eds.), Evolutionary Image Reliable Software Technologies - Ada-Europe'99. Pro-
Analysis, Signal Processing and Telecommunications. ceedings, 1999. XIII, 451 pages. 1999.
Proceedings, 1999. X, 225 pages. 1999. Vol. 1625: B. Reusch (Ed.), Computational Intelligence.
Vol. 1597: H. Zuidweg, M. Campolargo, J. Delgado, A. Proceedings, 1999. XIV, 710 pages. 1999.
Mullery (Eds.), Intelligence in Services and Networks. Vol. 1626: M. Jarke, A. Oberweis (Eds.), Advanced In-
Proceedings, 1999. XII, 552 pages. 1999. formation Systems Engineering. Proceedings, 1999. XIV,
Vol. 1598: R. Poll, P. Nordin, W.B. Langdon, T.C. Fogany 478 pages. 1999.
(Eds.), Genetic Programming. Proceedings, 1999. X, 283 Col. 1628: R. Guerraoui (Ed.), ECOOP'99 - Object-Ori-
pages. 1999. ented Programming. Proceedings, 1999. XIII, 529 pages.
Vol. 1599: T. Ishida(Ed.), Multiagent Platforms. Proceed- 1999.
ings, 1998. Vin, 187 pages. 1999. (Subseries LNAI). Vol. 1629: H. Leopold, N. Garcia (Eds.), Multimedia
Vol. 1601: J.-P. Katoen (Ed.), Formal Methods for Real- Applications, Services and Techniques - ECMAST'99.
Time and Probabilistic Systems. Proceedings, 1999. X, Proceedings, 1999. XV, 574 pages. 1999.
355 pages. 1999. Vol. 1634: S. Dzeroski, P. Flach (Eds.), Inductive Logic
Vol. 1602: A. Sivasubramaniam, M. Lauria (Eds.), Net- Programming. Proceedings, 1999. VIIL 303 pages. 1999.
work-Based Parallel Computing. Proceedings, 1999. VIII, (Subseries LNAI).
225 pages. 1999. Vol. 1639: S. Donatelli, J. Kleijn (Eds.), Application and
Theory of Petri Nets 1999. Proceedings, 1999. VIII, 425
pages. 1999.