IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO.

4, JULY/AUGUST 2024 1935

An Automated Multi-Layered Methodology to Assist

the Secure and Risk-Aware Design of Multi-Factor
Authentication Protocols
Marco Pernpruner , Roberto Carbone , Giada Sciarretta , and Silvio Ranise

Abstract—Authentication protocols represent the entry point Considering that authentication potentially allows to access
to online services, so they must be sturdily designed in order to sensitive information and perform restricted operations, mali-
allow only authorized users to access the underlying data. However, cious agents frequently target this phase: in the first quarter of
designing authentication protocols is a complex process: security
designers should carefully select the technologies to involve and 2022, almost 113 million attacks have been performed against
integrate them properly in order to prevent potential vulnerabil- multi-factor authentication systems, with the main objectives
ities. In addition, these choices are usually restricted by further being staffing/recruiting (4.45%), public (3.99%), and financial
factors, such as the requirements associated with the scenario, the (3.86%) services [2]. This trend is confirmed when we focus
regulatory framework, the dimensions to balance (e.g., security vs.
on financial services: 80% of the organizations have suffered
usability), and the standards to rely on. We come to the rescue
by presenting an automated multi-layered methodology we have from at least one cyber breach due to authentication weaknesses,
developed to assist security designers in this phase: by repeatedly and 95% of total breaches could be presumedly ascribable to
evaluating their protocols, they can select the security mitigations credential misuse or authentication vulnerabilities [3].
to consider until they reach the desired security level, thus enabling Given the considerable number of attacks, authentication
a security-by-design approach. For concreteness, we also show how
protocols should be designed in order to be sufficiently resistant
we have applied our methodology to a real use case scenario in the
context of a collaboration with the Italian Government Printing and guarantee a significant level of security. However, designing
Office and Mint. security protocols is not trivial: first of all, security designers
need to understand the scenario in which the authentication
Index Terms—Authentication, risk analysis, security analysis,
security methodology. protocol is deployed, as it usually imposes some constraints
regarding the environment and/or the legal framework to comply
with. Then, many other choices have to be made: the proper
I. INTRODUCTION balance between different dimensions (e.g., usability and pri-
vacy), the standards to rely on, and the authenticators to use in
HE National Institute of Standards and Technology (NIST)
T defines authentication protocols as «a defined sequence of
messages between a claimant and a verifier that demonstrates
the protocol, which must be properly configured and integrated
between each other without exposing the protocol to vulner-
abilities. During the process, security designers usually need
that the claimant has possession and control of one or more valid
to analyse the security of several configurations of the system
authenticators to establish their identity» [1]. Authenticators rep-
to identify the most appropriate one. However, performing a
resent the core elements within the authentication procedures,
manual analysis of each configuration would lead to greater
and can attest one or more authentication factors: something
efforts, longer development lifecycles and a higher probability of
that the claimant knows (knowledge factors), owns (ownership
missing some vulnerabilities [4], [5]. Automated techniques can
factors), or is (inherence factors).
greatly support the analysis process of authentication protocols
due to their reliance on advanced methods such as model check-
Manuscript received 19 October 2022; revised 20 April 2023; accepted 26 ers, which can turn the security analysis into large satisfiability
June 2023. Date of publication 17 July 2023; date of current version 11 July problems [6]. As a consequence, these techniques are fundamen-
2024. This work was supported by Futuro & Conoscenza Srl, jointly created
by Fondazione Bruno Kessler (FBK) and Poligrafico e Zecca dello Stato Ital- tal in detecting complex, uncovered vulnerabilities that affected
iano (IPZS, the Italian Government Printing Office and Mint). (Corresponding common security protocols [7]. Unfortunately, such techniques
author: Marco Pernpruner.) are computationally very intensive as they suffer the state space
Marco Pernpruner is with the Center for Cybersecurity, Fondazione Bruno
Kessler, 38123 Trento, Italy, and also with the Department of Informatics, explosion problem and may be difficult to exploit to quickly
Bioengineering, Robotics and System Engineering, University of Genoa, 16146 evaluate alternative configurations of a design.
Genoa, Italy (e-mail: [email protected]). To alleviate this problem and allow the usage of formal
Roberto Carbone and Giada Sciarretta are with the Center for Cybersecu-
rity, Fondazione Bruno Kessler, 38123 Trento, Italy (e-mail: [email protected]; techniques during the design phase, we have developed a multi-
[email protected]). layered methodology that supports security experts in the design
Silvio Ranise is with the Center for Cybersecurity, Fondazione Bruno Kessler, of authentication protocols, thus fostering a security-by-design
38123 Trento, Italy, and also with the Department of Mathematics, University
of Trento, 38123 Trento, Italy (e-mail: [email protected]). approach. Our methodology can be repeatedly employed as an
Digital Object Identifier 10.1109/TDSC.2023.3296210 oracle in the design process, with security designers giving in
© 2023 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see
https://creativecommons.org/licenses/by/4.0/
1936 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

Fig. 1. Graphical representation of our methodology.

input different configurations of the protocols until they reach violate the security goal G:
the desired security level. The multiple layers allow to satisfy
some requirements that we have identified from our experience A = {A ⊆ T M | MP , μ (A) |= G} ()
in protocol design and analysis: (i) Efficiency, to support security with the following specifications required as input:
experts in reasonable time; (ii) Classification, to understand the r a model of the protocol (MP ) derived from the Message
riskiness of the successful attackers. In particular, the combi- Sequence Chart (MSC, in blue), a detailed representation
natorial analysis performs a fast, yet incomplete, high-level of the protocol to be analysed from which we can extract
analysis to reduce the number of invocations of the symbolic the authentication factors used, the entities, their initial
analysis, which relies on formal frameworks and is connected knowledge, the messages they exchange, the communica-
with a higher computational complexity; this complies with the tion channels used, and the security assumptions;
first requirement. Finally, to meet the second requirement, the r a model of the attackers (MA ) obtained by using a list of
risk analysis complements the list of attackers that are able to potential attackers (i.e., a threat model T M) equipped with
violate the protocols with the related risk, to understand which a set of capabilities (in orange). In Section IV-B, we provide
of them represent the most relevant threats. a reference model that can be extended by a security
For the sake of concreteness, we also show how we have expert according to the needs. This model takes inspiration
applied our methodology to a real use case scenario: an au- from the Authenticator Threats introduced by NIST [8].
thentication procedure based on QR codes and electronic doc- To formalise the capabilities of a specific attacker, we
uments that currently represents one of the main authentica- introduce a function μ : T M → MA , which takes in input
tion procedures to access Italian Public Administration’s online an attacker A ∈ T M and returns the specification of the
services. This activity has been performed in the context of a corresponding capabilities from the model of the attackers
long-standing collaboration with Poligrafico e Zecca dello Stato MA .
Italiano (IPZS, the Italian Government Printing Office and Mint) In case of authentication protocols, G represents the fact that
and has allowed them to gradually refine the protocol until they the intended service (called Service Provider) must authenticate
reached the desired trade-off between security and usability. the user with a given level of assurance, which is specified by
the service itself depending on many factors.
A. Structure of the Paper From our experience in the design and analysis of authenti-
Section II describes the methodology that we have developed cation protocols, we have identified the following requirements
to support the design of authentication solutions. Section III that our methodology should meet:
introduces a concrete authentication protocol that we use as R1. Efficiency: the methodology should be efficient enough
a practical example. Section IV shows the application of the to allow security experts to promply obtain results during
methodology to the use case scenario. Section V identifies the design of authentication protocols;
and discusses the role of security mitigations along with their R2. Classification: the methodology should provide a clear
implications from security and usability perspectives. Section VI classification of the successful attackers to allow security
presents related works. Section VII draws some conclusions and designers to understand the related risk.
hints future works. For this reason, we have structured our methodology in
multiple layers: combinatorial analysis, symbolic analysis, and
II. SECURITY ANALYSIS METHODOLOGY risk analysis; this way, we can enable security designers to
make informed decisions while maintaining the cost of invoking
To support the design of authentication protocols, we have automated security analysis techniques at a reasonable level. In
developed a methodology – displayed in Fig. 1 – that aims at
detecting all the (combinations of) attackers A1 that are able to
own capabilities; in the second case, they need to collude and combine their
capabilities. For readability, single attackers A = {a} can be denoted as {a} or
1 It is worth underlining that A can represent either single attackers or combi- a; combinations of attackers A = {a1 , . . . , an } can be denoted as {a1 , . . . , an }
nations of attackers: in the first case, attackers are successful by relying on their or a1 +· · · +an .
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1937

particular, the combinatorial layer requires high-level specifica- TABLE I

ATTACKERS’ CAPABILITIES (MAC ) IN EXAMPLE 1
tions and performs a fast, yet incomplete, analysis whose role is
to reduce the number of attackers to test in the following layer –
which is more computationally expensive – and comply with R1.
To provide complete results, the second layer is represented by
the symbolic analysis, which relies on advanced frameworks
requiring formal specifications with cryptographic details; this
analysis suffers from the well-known state space explosion prob-
lem that is common to several state-based techniques [9]. Finally,
the risk analysis collects the results of the previous analyses and
complements them with a risk assessment that provides a basis
to plan the mitigations, thus complying with R2.
and possess the smartphone, thus violating the protocol (all the
padlocks are open in Table I).
A. Combinatorial Analysis Following Example 1, once we detect Thief + Social Engineer
The first layer of our methodology aims at finding: as successful, any wider combination involving these attackers
  would be trivially successful too, as attackers colluding together
AC = AC ⊆ T M | MPC , μC (AC ) |=C GC (C ) result in an enrichment of the original capabilities. As a conse-
quence, to avoid redundancy, we only consider minimum sets
with of attackers throughout our analyses: when a set of attackers
μ C : T M → MA C is detected as successful, we never consider any larger combi-
nation involving the already detected attacker. Formally, given
This is achieved through a high-level analysis based on au- a non-empty set AC ∈ AC , no combination AC   AC will be
thentication factors: GC holds when attackers are not able to considered.
compromise all the authentication factors involved in the pro- Given a successful violation, we observe that the following
tocol. In fact, attackers who compromise all the authentication two properties hold for the combinatorial analysis:
factors cause a violation of the whole protocol, and are reported r it never reports false positives, meaning that the attackers
by the combinatorial analysis. In this context, MPC represents detected by the analysis do manage to violate the secu-
the list of authentication factors involved in the protocol (inferred rity goal. Therefore, the combinatorial analysis is sound
from the MSC), while MAC represents the attackers’ capabilities w.r.t. (C );
in terms of compromised authentication factors: we can consider r it may miss some advanced attacks yielding to false nega-
them as a table where rows represent attackers and columns rep- tives, thus it is incomplete w.r.t. (C ).
resent authentication factors: each pair (a, af ) specifies whether These properties allow us to (dramatically) reduce the number
attacker a is able or not to compromise the authentication factor of invocations of the precise and resource-intensive security
af . We may also use the notation a af , where af can be a analysis in the next layer.
compromised authentication factor or ∅ in case the attacker a
has no effect on the protocol. B. Symbolic Analysis
In particular, AC is reported by the combinatorial analysis From the combinatorial analysis, we obtain the list of all the
(and thus is a solution of (C )) iff for each authentication explicit attackers violating the protocol (AC ). As a second layer
factor af ∈ MPC there exists an attacker a ∈ AC so that a of our methodology, we rely on the symbolic analysis that aims
af . These attackers, resulting from an explicit violation of the at finding:
authentication factors, are called explicit.
AS = {AS ⊆ T M \ AC | MPS , μS (AS ) |=S GS } (S )
Example 1. Let us consider an authentication protocol in
which users need to insert a password and scan a QR code with
through an application on their smartphone (on which they must
μ S : T M → MA S
already be authenticated). The protocol is composed of two
authentication factors: the password (i.e., a knowledge factor) The symbolic analysis can also discover complex attacks
and the smartphone (i.e., an ownership factor). Let us consider where the attackers do not need to compromise all the authentica-
two attackers: thieves, who physically steal devices, and social tion factors to violate the protocol, as they deceive the victim into
engineers, who deceive people into performing operations or re- implicitly compromising the remaining factors on their behalf;
vealing secrets; therefore, T M = {T hief, Social Engineer}. we have defined them as implicit attacks [10].
Thieves manage to steal the smartphone, but they do not know Example 2. Let us consider the protocol in Example 1. Social
the password, so they cannot compromise the protocol. Social engineers could launch an authentication request, insert the
engineers manage to know the password, but they do not possess user’s password (which they do know), and send the QR code to
the smartphone, so they cannot compromise the protocol. The the user, claiming that she might win a cruise by scanning it. As a
only way they have to compromise the protocol is by combining consequence, although they cannot explicitly compromise all the
their capabilities: together, they can both know the password authentication factors involved (they cannot possess the user’s
1938 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

smartphone, as per Table I), they anyway manage to implicitly Rating Methodology [15]: R computes the overall likelihood
compromise them as soon as the user scans the QR code. and impact by computing the average of the values returned by
By relying on the soundness of the combinatorial analysis, ρL and ρI , respectively; each of them is then assigned a label
we employ the symbolic analysis only to evaluate the attackers (Low, Medium, High). Finally, the likelihood and impact labels
who have not been detected by the combinatorial analysis (i.e., are combined according to a risk matrix (described in [15]) to
T M \ AC ); for this reason, the combinatorial analysis shares obtain the overall risk.
the list of successful explicit attackers. Below, for concreteness, we consider that ρL returns five
Given our expertise and past experience, for the symbolic factors and ρI four factors, as specified in Table II. We ob-
analysis we have chosen to leverage ASLan++ [11] – the spec- serve that the following discussion can be easily adapted to
ification language of the AVANTSSAR Platform [6] – in combi- other methodologies considering different tuples of factors to
nation with SATMC [9] – a model checker for security protocols characterise likelihood (ρL ) and impact (ρI ).
relying on advanced SAT solvers. Anyway, the protocol could Depending on the considered attacker A, we now distinguish
also be modelled in different formal specification languages two cases: A is a single attacker, i.e., |A| = 1; or A is a combi-
and given in input to other model checkers (e.g., ProVerif [12] nation of attackers, i.e., |A| > 1.
or Tamarin [13]). With respect to the combinatorial analysis, 1) Single Attacker (A = {a}): Given the set of factors that
here the inputs described in Section II have to be expressed in we have adopted:
more refined specifications based on ASLan++. As a conse-
ρL (A) = ρL (a) = vTa D , vOa , vAV
a
, vUa I , vSA
a
quence, MPS does not represent just the authentication factors
a a a a
involved in the protocol, but consists in a refined model of ρI (A) = ρI (a) = vLSP , vAS , vAD , vAP
the messages exchanged by the entities. GS is modelled by
specifying the security properties that must hold on the channel where vfa = [0, 9] corresponds to the value assigned to the factor
that gets established between the user and the Service Provider. f for the attacker a.
Moreover, while MAC was specified in terms of compromised Finally, the risk can be computed by combining the likelihood
authentication factors, MAS is modelled in terms of channels’ and impact through a suitable function R:
properties that get violated and knowledge that is acquired dur- Risk (A) = Risk(a) = R (ρL (a), ρI (a))
ing the protocol. The Dolev-Yao model (MDY ) [14] supported
in SATMC can thus be extended by specifying custom attackers 2) Combination of Attackers: (A = {a1 , . . . , an }): When
with additional capabilities (e.g., those belonging to our threat considering a combination of attackers, each ak ∈ A (with
model), so that MAS = (MDY  MA∗ S ). k = 1, . . . , n) is associated with a set of likelihood and impact
Example 3. From Example 2, social engineers can threaten values:
the confidentiality between the user and the mobile application ρL (ak ) = vTakD , vOak , vAV
ak
, vUakI , vSA
ak
(and viceversa), as they can deceive the user into revealing
ak a1 ak ak
whichever value she inserts in (or reads from) the application. ρI (ak ) = vLSP , vAS , vAD , vAP
Moreover, they can compromise the authenticity between the
As a consequence, for each likelihood and impact factor,
browser and the user, as the QR code that she scans would not
we have n values. We now explain how to derive a single
really come from the browser (even though the user may think
tuple for likelihood and one for impact from the n available
so).
tuples, respectively, so to apply the function R to derive the risk
Once the model checker receives MPS and MAS , it verifies
value. For this, we define a function C taking as input n values
that they satisfy the security goal, reporting details about the
belonging to a given factor f :
attack in case a violation is found. We observe that the symbolic  
analysis is both sound and complete with respect to (S ) under VfA = C vfa1 , . . . , vfan
suitable assumptions.
With respect to the case of the single attacker, the functions
ρL and ρI have to be redefined accordingly:
C. Risk Analysis
At the end of the combinatorial and symbolic analyses, a ρL (A) = VTAD , VOA , VAV
A
, VUAI , VSA
A

list of explicit (AC ) and implicit (AS ) attackers is generated A A A A

ρI (A) = VLSP , VAS , VAD , VAP
and shared with the risk analysis layer, which evaluates the
risks of each attacker as a combination between the likelihood Finally, the risk can be computed as follows:
and impact, according to some factors. Let ρL and ρI be two
Risk (A) = R (ρL (A) , ρI (A))
functions that, given an attacker in T M, return tuples containing
parameters that once combined together yield the likelihood In our analysis, we have defined C as follows:
or impact values, respectively; these values are defined by the r In general, we consider the minimum between all the
security expert for each single attacker. Moreover, let R be values for the considered factor. For instance, a combi-
a function that computes the risk, by suitably combining the nation of a physical thief (which must act physically, thus
likelihood and impact. Although different methodologies can vAV = 1) and a remote malware (which can act remotely,
be used, we rely on an extended version of the OWASP Risk thus vAV = 9) needs a physical intervention anyway, hence
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1939

TABLE II
FACTORS CONSIDERED FOR THE RISK ANALYSIS

the former value will be considered. Considering f ∈ of the attack. However, this consideration only applies
{T D, AV, U I, LSP, AS, AP }: when more than a single device is stolen. Therefore, given
    a combination of attackers that manage to physically steal
C vfa1 , . . . , vfan = min vfa1 , . . . , vfan k devices:
r Beyond the previous consideration, when considering  a an

C vAD1
, . . . , vAD
some particular factors such as the Opportunity and the
 a1 an
Spread of Attack, the likelihood decreases as the number of min vAD , . . . , vAD if k = 0
attackers involved n increases. Considering f ∈ {O, SA}: = a1 an
min vAD , . . . , vAD − (k − 1) if k ≥ 1
   
C vfa1 , . . . , vfan = min vfa1 , . . . , vfan − (n − 1) Remind that, in any case, values range from 0 to 9.
r In addition, we can distinguish two classes of attackers:
physical and remote (according to how they perform the D. Relationship Between the Analyses
attack). When attackers from both the classes are involved
in a combination, they need not only to perform physical Our methodology involves two different layers for the security
operations, but also to act remotely on the same user’s analysis (i.e., the combinatorial and symbolic analyses), which
devices, thus further reducing both the Opportunity and aim at identifying the explicit and implicit attackers that are able
the Spread of Attack. Considering f ∈ {O, SA}: to compromise the considered protocol. The list of successful
    attackers is then shared with the risk analysis layer, which
C vfa1 , . . . , vfan = min vfa1 , . . . , vfan − (n − 1) − 2 assigns a risk value to each of the attackers detected by the
previous analyses.
r Finally, as far as the Attack Detection is concerned, the The combinatorial and symbolic analyses are connected by
number of physical devices stolen must be taken in consid- a strong relationship that must be taken in consideration when
eration, since a higher number results in an easier detection providing the related models: MAS extends MAC , i.e., MAS
1940 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

models in the specification language at least the same capabili- the corporate sector, to support the design of new authentication
ties on the authentication factors represented by MAC . To ensure schemes or analyse existing ones according to internal needs.
consistency between the two analyses, these models should be
provided by a security expert. III. USE CASE SCENARIO
Due to this relationship, the combinatorial analysis is not
We have a long-standing collaboration with Poligrafico e
strictly necessary, as the symbolic analysis would be able to de-
Zecca dello Stato Italiano (IPZS), the Italian Government Print-
tect both explicit and implicit attackers by itself. In particular, for
ing Office and Mint that is responsible for producing the Italian
each attacker AC detected by the combinatorial analysis, there
eID card, called Carta d’Identità Elettronica (CIE 3.0) [17].
exists a corresponding attacker AS detected by the symbolic
This collaboration aims at developing and analysing cutting-
analysis, with AC = AS ; therefore, the combinatorial analysis
edge identity management solutions to fully benefit from the
is sound with respect to the symbolic analysis. Consequently, we
capabilities of eID cards, which are equipped with a microchip
rely on the combinatorial analysis to reduce the set of attackers
that communicates through a contactless NFC interface. They
to test in the symbolic analysis, so as to reduce the overall
are also provided with an X.509 certificate [18] containing the
complexity of the process.
personal data of the owner, whose trustworthiness is ensured
On the contrary, the symbolic analysis is mandatory: for each
by the competent authorities through a digital signature. As a
attacker AS , there not always exist a corresponding attacker
consequence, each card has a pair of keys that can be used
AC , with AS = AC ; therefore, the combinatorial analysis is
for public-key cryptography (following the IAS ECC specifi-
incomplete with respect to the symbolic analysis.
cations [19]), whose use can be unlocked by inserting a PIN.
Involving eID cards in an authentication workflow may pro-
E. Computational Considerations vide several benefits from a security perspective. In this context,
The symbolic analysis requires advanced computational capa- we focused on hybrid scenarios allowing users to authenticate
bilities, which result in a higher time of execution. On the other from a personal computer by leveraging their smartphone as
hand, it provides a higher level of confidence on the results, an eID card reader. In particular, we now describe a concrete
meaning that it reports also complex attacks that may have been protocol that is currently used in the Italian ecosystem as one
missed during the combinatorial analysis. If we performed the of the main authentication procedures to access Public Admin-
symbolic analysis for all the n attackers (and combinations) istration’s online services. This protocol requires users who
belonging to the threat model, we would have to run it 2n − 1 wish to authenticate themselves to launch a request from their
times. We can considerably reduce the set of attackers to test personal computer’s browser, which then displays a QR code.
thanks to the following considerations: Once users scan the QR code via a custom mobile application
C1. Explicit attacks from the combinatorial analysis: the (namely, eIDApp) on their smartphone, they are guided through
main goal of the combinatorial analysis is to prune the set the interaction with their eID card and finally authenticated on
of attackers to test, given its speed in detecting explicit their personal computer. Therefore, the authentication factors
attackers. Consequently, by relying on its soundness, involved are the eID card ( ) and the PIN ( ).
we check neither already detected attackers nor larger
combinations involving sets of successful attackers. A. Entities
C2. Physical thieves: when they steal a device, users cannot The protocol involves the following set of entities:
be deceived into approving a malicious authentication r User: the claimant, who wants to authenticate on a specific
attempt on that device, because they no longer own it. service. Each user is supposed to have a userId, which
Therefore, these attackers cannot be involved in implicit is a uniquely identifying value that is contained in the eID
attacks. card’s certificate.2
C3. Non-minimum combinations of implicit attacks: once r eID card: the eID card belonging to the user.
the symbolic analysis detects successful attackers, we r eIDApp: the mobile application responsible for securely
do not consider further combinations with them (see interacting with the eID card.
Section II-A for a formal definition). r Identity Provider (IdP): responsible for managing users’
More details will be provided in Section IV-E. online identities and ensuring proper authentication. It is
composed of a front-end interface (accessible through a
F. Application Scenarios browser) and a back-end server (IdPServer).
In the next section, we describe a concrete authentication
r Service Provider (SP): any online service where users can
protocol to which we apply our methodology. This use case authenticate. It is composed of a front-end interface (acces-
scenario just aims to provide a concrete example of how our sible through a browser) and a back-end server (SPServer).
methodology can be applied and how the results can be used
r Browser: a web browser that users can interact with in order
to support the design phase. In fact, our methodology can be to access front-end interfaces of both the SP and the IdP.
employed to analyse authentication protocols in many more This browser runs on a personal computer belonging to the
scenarios. For example, in the financial sector, to highlight the user.
differences between online banking authentication protocols be-
fore and after the PSD2 regulation on payment services [16]; in 2 In the Italian scenario, the serial number of the eID card is used as userId.
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1941

Fig. 2. Message sequence chart of the protocol.

The IdPServer and the SPServer are part of a trust relationship, 2) User info: the user fills her identifying information
obtained after a federation procedure [20]. (userId) in the login page. The browser retrieves this
value and sends it to the IdPServer.
3) Challenge: the IdPServer generates the challenge and
B. Flow displays it on the browser as a QR code, which the user is
Fig. 2 shows the message sequence chart of the protocol, required to scan through the eIDApp to extract the plain
which is composed of the following phases: challenge.
1) Authentication request: the user visits the SP webpage 4) Response: after providing the PIN, the user is required
and launches a new authentication request, thus being to place her eID card near the mobile device for NFC
redirected to the IdP. The IdPServer generates some scanning. In case the PIN is correct and has been provided
fresh values associated with the authentication attempt: within a fixed number of attempts, the eID card signs
an identifier opId and an associated cookie IdPSes- the challenge through its private key, thus generating
sionCookie that will be used during any commu- the response that is finally sent back to the IdPServer
nication between the browser and the IdPServer. After through the eIDApp along with the eID card’s certificate.
storing these values, the IdPServer displays the login 5) OTP verification: the IdPServer generates a fresh OTP
page. and associates it with the current opId. The OTP, which
1942 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

TABLE III
SECURITY ASSUMPTIONS

is displayed to the user through the eIDApp, needs to be IV. APPLICATION OF THE METHODOLOGY
written back to the IdP login page on the browser. Finally, In this section, we show how we have applied our methodol-
it is sent to the IdPServer.
ogy to the use case scenario defined in Section III.
6) Redirection: the IdPServer redirects the user to the
SPServer with an authentication assertion.
A. Security Setup
During the protocol, the IdPServer verifies that:
r Step 10: the incoming IdPSessionCookie matches the In this section, we set the analyses up for our use case scenario
one generated at step 5. by defining the security assumptions (Section IV-A1) and the
r Step 23: (i) the eID card’s certificate has not been revoked; attackers’ capabilities (Section IV-A2).
(ii) the userId previously inserted matches the one stored 1) Security Assumptions: During the analyses, we consider
in the eID card’s certificate; (iii) the value obtained after the security assumptions that are described in Table III. We
applying the eID card’s public key to response matches categorize them in: eID cards Assumptions (EA), ensuring
the challenge. that the authenticator has been properly activated; Procedural
r Step 29: (i) the incoming IdPSessionCookie matches Assumptions (PA), dealing with the authentication procedure
the one generated at step 5; (ii) the OTP inserted by itself; Trust Assumptions (TA), related to the trust between the
the user on the browser at step 27 matches the one entities involved in the protocol; and Channels Assumptions
generated by the IdPServer at step 24; (iii) the OTP (CA), regarding the properties of the communication channels.
has been provided within a fixed number of attempts; 2) Threat Model and Attackers’ Capabilities: In order to
(iv) the operation has been completed in a fixed time analyze the security of the protocol, we have identified from [8]
interval. the Authenticator Threats that can violate it. Then, we have
In case one of these checks fails, the authentication procedure expanded and contextualized the related attackers to obtain the
ends with an error. threat model in Table IV:
T M = {PCT, MDT, CT, D, ES, SS, SE, MB, MM}
C. Challenge The relationship between the Authentication Threats identi-
During authentication, the IdPServer needs to be sure that the fied by NIST and our threat model is detailed in the comple-
involved eID card really belongs to the user who is going to be mentary website [21]. In particular, we have not considered:
authenticated. To this end, a challenge–response procedure [1] “Assertion Manufacture or Modification” (since the authentica-
occurs between the IdPServer and the eID card. In the consid- tion assertion is digitally signed by the IdPServer and cannot be
ered protocol, the challenge is composed of the following tampered with); “Offline Cracking” and “Online Guessing” (due
parameters: to the restricted number of possible attempts while inserting the
r opId: the operation identifier, which is randomly gener- eID card’s PIN); “Side Channel Attack” (as in EA4 we assume
ated and represents the authentication attempt; that the eID card’s private key is particularly difficult to extract);
r userId: the user identifier, which the user fills in during and “Unauthorized Binding” (as eID cards can be associated
the procedure; only to their legitimate owners, due to EA1).
r IdPName: the name of the IdP used for the authentication; Once defined the attackers, we have also defined their capabil-
r SPName: the name of the SP which the user wishes to ities in terms of compromised authentication factor(s) in our sce-
authenticate onto; narios (Table V): closed padlocks ( ) denote non-compromised
r opText: a textual description of the current operation, factors, while open padlocks ( ) represent compromised fac-
displayed on the mobile device before the approval. tors. In addition, we use an asterisk (  ) to indicate a possession
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1943

TABLE IV
ATTACKERS BELONGING TO THE THREAT MODEL

TABLE V in ASLan++ following the work in [22]. To better understand

ATTACKERS’ CAPABILITIES IN THE COMBINATORIAL MODEL
the formalisation, Table VI shows some relevant predicates in
ASLan++. The complete ASLan++ models are available on
the complementary website [21].
1) Protocol (MPS ): The model of the protocol formally de-
scribes the entities involved and the communications between
them over the communication channels. This model should be as
consistent as possible with the protocol it describes, in order to
obtain a proper analysis; however, as often happens, models can
also contain some approximations to reduce the computational
complexity of the analysis, though without losing crucial details.
factor that is compromised indirectly. For instance, a malicious For space reasons, the approximations adopted in our models are
application does not physically violate eID cards, yet it manages described on the complementary website [21].
to deceive victims into interacting with their eID cards, thus Legitimate entities may be required to know specific infor-
compromising that factor without physically possessing it. mation before the protocol execution. ASLan++ requires these
values to be passed as arguments to the related entities. In the
B. Combinatorial Analysis considered protocol:
r the user knows the PIN of the eID card and the userId;
The combinatorial analysis, described in Section II-A, dis- r the eID card knows its PIN.
covered 5 attackers that are able to compromise the protocol
explicitly: In addition to honest entities, also attackers (which in
ASLan++ are referred to as intruders and indicated with i)
AC = {{MM} , {CT, D} , {CT, ES} , {CT, SS} , {CT, SE}} may have some preliminary knowledge. In ASLan++, an initial
knowledge k is given to the intruder by using the expression
where the first corresponds to a malicious application, while the iknows(k). However, since attackers’ initial knowledge de-
others correspond to a card thief combined with another attacker pends on their capabilities, it will be specified directly in the
able to discover the user’s eID card’s PIN. attackers’ model MA∗ S .
For instance, considering the combination CT+SS: CT man- Beyond modelling the protocols, we also had to formalise
ages to obtain the victim’s eID card ( ) by physically stealing the security assumptions that we have identified in Table III.
it, while SS can compromise the PIN of the eID card ( ) by The formal specification of such assumptions can be found
looking at the victim while typing it. As a consequence, the in Table VII, while the description of the predicates used is
combination violates all the authentication factors involved in provided in Table VI.
the QR protocol and thus is able to authenticate onto an online 2) Model of the Attackers’ Capabilities (MA∗ S ): To model
service as the victim. the attackers’ capabilities for the symbolic analysis, we modify
some parts of the specifications according to the attacker that
C. Symbolic Analysis we are considering in the specific run of the symbolic analysis.
As explained in Section II-B, we need to provide SATMC Below you can find a description of how each attacker affects the
with all the parameters in input, which have been modelled protocol; the numbers in parentheses refer to Table VIII, which
1944 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

TABLE VI
RELEVANT PREDICATES IN ASLAN++

TABLE VII
SECURITY ASSUMPTIONS IN THE SYMBOLIC MODEL

displays the corresponding changes to make in the ASLan++ r since the attacker is the only entity who can interact with
model. We also show the relationship between the symbolic the browser (4), the eIDApp (8) or the eID card (12), he
and the combinatorial models of the attackers’ capabilities, to uses the same communication channel in every session.
underline that the former (MAS ) can always be reduced to the This corresponds to violating the personal computer, mobile
latter (MAC ); this leads to the discussion in Section II-D. device, or eID card, respectively. However, in our protocol,
It is important to notice that we provide instructions to model only the eID card is considered as an authentication factor
both how to consider and how not to consider a specific attacker. (P CT, M DT ∅; CT ).
Therefore, a dash (—) means that the predicate in the other D The attacker manages to copy the eID card’s PIN that may be
column of the same row does not apply.
written on paper (D ), thus getting to know this value
(13).
PCT, MDT, CT Before the attacker steals the user’s device ES, SS Without considering these attackers, the following in-
(personal computer, mobile device or eID card), every inter- formation can be known only to the intended recipient, thus
action with it is surely made by the user herself, thus the the considered communication channels are confidential:
channel between the user and the browser (1), the eIDApp
(5) or the eID card (9) is authentic. Moreover, the fact r what the user types in the eIDApp/browser can be known
userOwnComputer (2), userOwnSmartphone (6) or only by the eIDApp (14)/browser (15);
userOwnEIC (10) needs to be accordingly set to true, r what the browser shows to the user can be known only by
since the user physically owns its device. the user (16).
On the contrary, when the user’s device is possessed by the Property (15) corresponds to violating the PIN, which the user is
attacker, all the interactions are made by the same entity (the required to insert in the eIDApp (ES, SS ). Instead, prop-
attacker himself). Therefore: erties (15) and (16) are needed because users are displayed an
r since the browser (3), the eIDApp (7) and the eID card OTP on the mobile application and have to insert this OTP on the
(11), respectively, do not have any guarantee on this entity’s personal computer’s browser, thus the corresponding channels
identity, the related channel is only weakly authentic; need to be confidential when the attacker is not considered.
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1945

TABLE VIII
ATTACKERS’ CAPABILITIES IN THE SYMBOLIC MODEL

All the properties above are no longer true when the attackers MM This attacker can take full control of the user’s mobile
become effective, since they manage to intercept all these values. device and perform any operation he wishes. Therefore, we
SE Without considering this attacker, the communications be- model MM by explicitly making the attacker impersonate the
tween the user and the browser are protected by the following eIDApp (24). However, we need to restrict other attackers’
properties: capabilities when the MM is not to be considered, otherwise
r what the user types in the browser can be known only by they could be too powerful than how we have really modelled
the browser, thus the channel is confidential (17); them. To this end, when MM is not considered:
r the user is sure that whatever the browser shows to her r the user is sure that whatever the eIDApp shows to her
really comes from the original browser, thus the channel is really comes from the original eIDApp, thus the channel is
authentic (18). authentic (22);
Properties (17) and (18) are needed because users are displayed r the eID card is sure that whatever the eIDApp sends to her
a QR code on the personal computer’s browser and at the end really comes from the original eIDApp, thus the channel is
of the operation they have to insert an OTP on the personal authentic (23).
computer’s browser, thus the corresponding channels need to be When considering MM, there is no need to remove instructions
confidential and authentic, respectively, when the attacker is not (22) and (23), as the channels do remain either authentic or
considered. confidential. However, by impersonating the eIDApp due to (24),
When SE becomes effective, instead, these properties are no MM can deceive the user into interacting with her eID card
longer valid: the attacker can both deceive users into revealing
(M M ), as well as know the PIN when the user types it
what they typed in the browser, and provide them with some
malicious values by pretending to be the browser (e.g., the (M M ).
QR code containing the challenge). Moreover, the attacker 3) Security Goal (GS ): In ASLan++, a channel goal has the
manages to make the user reveal the PIN of the eID card (19) following form:
and the OTP generated by the eIDApp (20), thus compromising name:(_) Sender Channel Receiver
We can rely on this syntax to model the security goal (identi-
these values (SE ).
fied in Section II) as follows:
MB This attacker can take full control of the user’s browser User_authn_to_SP:(_) User *->> SPServer;
and perform any operation he wishes, thus we model MB by This represents a goal called User_authn_to_SP that
making the attacker impersonate the browser (21). However, must be satisfied in the run(s) of the protocol. Specifically,
it does not violate any authentication factor in our protocol the goal requires that a communication channel between the
(M B ∅). user (sender) and the SPServer (receiver) gets established at the
1946 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

Fig. 3. Trace of the implicit attack performed by MB.

end of the protocol; on this channel, the following properties process on the same SP by using the same userId and obtain a
(represented by *->>) must hold: challenge in the form of a QR code. Following the authentication
r authenticity: guarantees that any incoming message on this process, the victim should be displayed a QR code as well, but
channel indeed comes from the user; – since the attacker has full control of the victim’s browser – he
r directedness: guarantees that any incoming message on can tamper with the victim’s authentication page and replace the
this channel was indeed intended for the SPServer; original QR code with that obtained in his parallel authentication
r freshness (or replay protection): guarantees that any mes- session. The user will be deceived into scanning the malicious
sage sent on this channel can be received only once. QR code with the eIDApp, inserting the PIN and reading the eID
4) Number of Sessions: For each analysis, we run two paral- card through NFC. At the end, she will be displayed an OTP
lel sessions of the protocol. This way, we can evaluate an attacker on the eIDApp to insert on her personal computer’s browser;
leveraging a parallel session launched by the users themselves however, having control of the browser, MB can intercept the
to finalise the attack, which represents how implicit attacks are OTP and insert it in his own authentication page. As a result,
usually performed. The two sessions share the same inputs, the attacker will be authenticated on his own personal computer
though the communication channels used are different (except with the victim’s credentials, as the QR code scanned by the
when modelling some attackers, as detailed above). user had been originally issued in the context of the attacker’s
5) Results: The symbolic analysis tested the following at- authentication session.
tackers: D, ES, SS, SE, MB, D+ES, D+SS, ES+SS, D+ES+SS The second implicit attack, performed by SE, is similar to the
(see Section IV-E for more details) and detected two attackers: previous one. However, SE does not need to alter the victim’s
browser, since he can provide her with the QR code through
AS = {{SE} , {MB}}
other means (e.g., via email or a social media) and deceive her
For both of them, the analysis reported the attack trace,3 which into scanning it through social engineering techniques (e.g., «S-
is a graphical representation of the messages exchanged between can the QR code and use your eID card to win a wonderful
the entities taking part in the protocol, by using arrows labelled cruise!»). Then, the attacker can – again – deceive the victim
with the content of the message. Messages consisting in the into revealing the OTP, so that he can finalise the authentication
concatenation of more values are joined by a dot, while fresh by impersonating the user.
values f are represented by the expression n(f ). In addition
to the entities taking part in the protocol, attack traces usually
D. Risk Analysis
display an additional entity i representing the intruder (i.e., the
attacker). In case the attacker impersonates another entity e, Table IX shows the results of the risk analysis applied to
this fact is represented by the expression i(e) placed inside our use case scenario, where we can identify three attackers
a box. associated with a low risk (MM, CT+ES and CT+SE), three
Considering MB, the attack trace in Fig. 3 shows that when the with a medium risk (CT+D, CT+SS and SE) and one with
victim tries to authenticate on an SP (through request1) and a high risk (MB). MB is a powerful attacker that requires
inserts her userId, the attacker can initiate an authentication a specific technical preparation to infect the browser, can be
performed fully remotely, does not need any interaction with
3 To enhance readability, the attack traces displayed in this paper are a the user, is difficult to detect, and can perform large-scale
simplified version of those generated by SATMC. attacks.
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1947

TABLE IX
RESULTS OF THE RISK ANALYSIS

TABLE X M2: Restricts the attack surface: attackers cannot just send
COMPUTATIONAL EFFECTS OF OUR CONSIDERATIONS DURING OUR ANALYSES
malicious QR codes to random people, but they need
to choose a precise victim as the userId is part of
the challenge. Although we do not require the userId
to be a secret value (like a password), it should not
be commonly known (like the name of the user) to
reduce the possibility of a general attack. Given the
additional value users have to insert, this mitigation
slightly reduces usability.
M3: Restricts the ability of attackers to deceive users by
sending improper QR codes, since the operation must
be completed within a certain time interval. This mit-
E. Computational Considerations igation partially affects usability as well, considering
In Section II-E we have proposed some considerations to that expired authentication attempts have to be launched
reduce the computational complexity of the symbolic analysis. again.
Table X displays the number of attackers that our considerations M4: By restricting the possible attempts, prevents guessing
exempted us from testing in the use case. When applied in se- and brute-force attacks on secret values. The number of
quence, each consideration further improves the set of attackers available attempt should be carefully set in order to find
excluded by the previous ones. As a result, the symbolic analysis a trade-off between security and usability (i.e., neither
needed to test only 9 attackers in our use case scenario (1.8%), too tight, nor too loose).
thus considerably optimising the analysis flow. M5: Restricts the possibility of an attack, since the attacker
would also need to obtain the OTP associated with that
specific authentication attempt. Specifically, we have
V. SECURITY MITIGATIONS decided to display the OTP on the mobile application in
The role of security mitigations is extremely important to order to reduce phishing attacks carried out via email:
reduce risks: they play a fundamental role in shortening the in case users do not have an authentication attempt
list of successful attackers or reducing the likelihood and/or currently ongoing on the personal computer’s browser,
impact of certain attackers. As a consequence, the selection of they would not know where to insert the OTP. Given
which mitigations are worth implementing is a crucial phase the additional value users have to insert, this mitigation
during protocol design. However, companies or governmental slightly reduces usability.
agencies implementing authentication procedures could have M6: Reminds users of verifying that QR codes are displayed
custom requirements; that imposes trade-offs between usability on an official website, thus reducing the likelihood of
and security when choosing mitigations. Table XI displays the attackers sending improper QR codes on phishing web-
mitigations that are implemented in our use case scenario. In sites (such as SE and MB). However, since the warning
addition to security, we focus on usability that is one of the most is always displayed within the mobile application, users
important dimensions to consider when selecting mitigations, could ignore or get used to it. This mitigation has no
as users are more willing to accept simple protocols rather than impact on usability, as it does not restrict users’ attempt
cumbersome ones. We now discuss each mitigation along with nor it requires users to perform additional operations.
their effects. M7: Helps the user distinguish between legitimate and ma-
M1: Since rooted devices are known to be extremely vulnera- licious authentication attempts. This mitigation is ex-
ble to common attacks, preventing the use of the eIDApp tremely effective in authorization contexts, as precise
on such devices brings many advantages in terms of details about the ongoing operation to authorize are
security. On the other hand, it may result in usability displayed to the user. When dealing with authentication,
issues since people who rooted their devices on purpose instead, it is difficult to find suitable details to uniquely
would not manage to use the application. identify the ongoing attempt, thus the security benefits
1948 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

TABLE XI
LIST OF POSSIBLE MITIGATIONS

TABLE XII
RESULTS OF THE RISK ANALYSIS WITH ONLY M4 APPLIED

are slightly lower. In general, the information displayed of the mitigations aim at targetting these attackers. As a re-
to the user should be relevant and uniquely identify sult, the risk of these powerful attackers can be significantly
the operation, otherwise some attacks could anyway be reduced.
performed. The information can be shown either without
affecting the procedure (e.g., in the same window where VI. RELATED WORK
the user inserts the PIN) or by introducing an ad hoc
activity, which would slightly affect usability. When The scientific literature contains many approaches to the
rating usability in Table XI, we consider the former case. security analysis of authentication protocols. For the sake of
brevity, we discuss only the approaches that are more relevant
for our work.
A first approach consists in proposing a new au-
thentication scheme and performing a formal analysis to
A. What if...? demonstrate its compliance with a given set of require-
In Section IV, we have analysed the use case scenario that ments and security goals. Many scientific works follow
implements all the mitigations listed in Table XI. However, it this approach in different contexts: healthcare [23], [24],
may be interesting to understand the effects of removing some generic [25] and industrial [26], [27] IoT environments,
mitigations, e.g., to improve the usability level of the designed smart homes [28], [29], wireless sensor networks [30], [31],
protocol. To this end, let us now consider the protocol in Fig. 2 and many more. However, several security analyses have
only implementing mitigation M4, which is already enforced by been found flawed [32], thus leading to potentially incorrect
eID cards themselves (as far as their PIN is concerned). Table XII results.
displays the results of the new risk analysis, highlighting a signif- Another approach aims at analysing the security of existing
icant worsening of the situation: SE becomes a critical attacker, authentication protocols or standards, such as FIDO [33], [34],
MM is now associated with a high risk, and the combination OAuth2.0 [35], 5G EAP-TLS [36] and Single Sign-On [22]. In
CT+SE increases to a medium risk. These results show that this case, the analysis aims at validating a third-party authenti-
carefully selecting the mitigations to implement during the de- cation scheme that is widely used.
sign phase brings many improvements to the security level of the In general, most of the analyses rely on formal frameworks,
protocol. which requires security experts to model the considered sce-
In general, SE, MB and MM are clearly the most powerful nario; then, they use formal provers or model checkers (such as
attackers, as they can deceive users through social engineer- ProVerif [12] or Tamarin [13]) to assess the security of the pro-
ing techniques or compromise users’ devices. Therefore, most tocol, finally obtaining a list of the attackers that can violate the
PERNPRUNER et al.: AUTOMATED MULTI-LAYERED METHODOLOGY TO ASSIST THE SECURE AND RISK-AWARE DESIGN 1949

security goals. These techniques usually analyse all the attackers ACKNOWLEDGMENT
contained in the threat model, thus resulting in a high execu-
The authors would like to thank the anonymous reviewers for
tion time associated with the computational complexity of the their valuable comments that helped us improve the quality of
process.
the article.
Moreover, the mere list of successful attackers may not be
enough: in a corporate scenario, for instance, security designers
could need to have clear indications on the risks associated
with the successful attackers, in order to prioritise them and REFERENCES
understand which ones need to be mitigated more urgently and [1] P. A. Grassi, M. E. Garcia, and J. L. Fenton, “Digital iden-
which of them can be ignored. tity guidelines,” NIST Special Publication 800-63-3, Jun. 2017,
By combining different level of analysis, our methodology doi: 10.6028/NIST.SP.800-63-3.
[2] Auth0, The State of Secure Identity 2022, 2022. [Online]. Available: https:
provides several benefits: the combinatorial analysis can be //auth0.com/resources/whitepapers/2022-state-of-secure-identity-report
performed even by less-expert users, and provides a list of [3] Vanson Bourne, The State of Authentication in the Finance Indus-
explicit attacks; the symbolic analysis still needs to be set up by try 2022, Jul. 2022. [Online]. Available: https://get.hypr.com/state-of-
authentication-in-the-finance-industry-2022
security experts, but is performed on a smaller set of attackers as [4] S. An et al., “CloudSafe: A tool for an automated security analysis for cloud
it only searches for implicit attacks, thus reducing the computa- computing,” in Proc. IEEE 18th Int. Conf. Trust Secur. Privacy Comput.
tional complexity of the analysis; the risk analysis complements Commun./IEEE 13th Int. Conf. Big Data Sci. Eng., 2019, pp. 602–609.
[5] G. Agosta, A. Barenghi, A. Parata, and G. Pelosi, “Automated security
the list of attackers by associating the related risks. Security analysis of dynamic web applications through symbolic code execution,”
designers are thus provided with a methodology that they can in Proc. 9th Int. Conf. Inf. Technol. New Gener., 2012, pp. 189–194.
customise according to their needs (e.g., they can skip the sym- [6] A. Armando et al., “The AVANTSSAR platform for the automated val-
idation of trust and security of service-oriented architectures,” in Proc.
bolic analysis if they need for quick, yet possibly incomplete, Int. Conf. Tools Algorithms Construction Anal. Syst., Springer, 2012,
results). pp. 267–282.
[7] A. Armando, R. Carbone, L. Compagna, J. Cuellar, and L. Tobarra,
“Formal analysis of SAML 2.0 web browser single sign-on: Breaking the
SAML-based single sign-on for Google apps,” in Proc. 6th ACM Workshop
Formal Methods Secur. Eng., New York, NY, USA, 2008, pp. 1–10.
VII. CONCLUSION [8] P. A. Grassi et al., “Digital identity guidelines: Authentication and lifecy-
cle management,” NIST Special Publication 800–63B, NIST, Jun. 2017,
In this paper, we have presented a multi-layered security doi: 10.6028/NIST.SP.800-63b.
methodology to analyse multi-factor authentication protocols. In [9] A. Armando, R. Carbone, and L. Compagna, “SATMC: A SAT-based
model checker for security protocols, business processes, and security
addition to identifying the list of attackers that are able to com- APIs,” Int. J. Softw. Tools Technol. Transfer, vol. 18, no. 2, pp. 187–204,
promise the protocol, our methodology provides information Apr. 2016.
about the associated risks. For concreteness, we have showed [10] M. Pernpruner, R. Carbone, S. Ranise, and G. Sciarretta, “The good, the
bad and the (not so) ugly of out-of-band authentication with eID cards and
how we applied the methodology to a real use case scenario: an push notifications: Design, formal and risk analysis,” in Proc. 10th ACM
authentication procedure based on QR codes and electronic doc- Conf. Data Appl. Secur. Privacy, 2020, pp. 223–234.
uments that currently represents one of the main authentication [11] AVANTSSAR, “ASLan++ Specification and Tutorial. Deliverable
D2.3”, Mar. 2011. [Online]. Available: https://st.fbk.eu/complementary/
procedures to access Italian Public Administration’s online ser- TDSC2022
vices. This activity, performed in the context of a long-standing [12] B. Blanchet, B. Smyth, V. Cheval, and M. Sylvestre, “ProVerif 2.00:
collaboration with Poligrafico e Zecca dello Stato Italiano (the Automatic Cryptographic Protocol Verifier, User Manual and Tutorial”,
May 2018. [Online]. Available: https://prosecco.gforge.inria.fr/personal/
Italian Government Printing Office and Mint), supported the bblanche/proverif/manual.pdf
design of the authentication protocol by highlighting which [13] S. Meier, B. Schmidt, C. Cremers, and D. A. Basin, “The TAMARIN
mitigations reached the best trade-off between security and prover for the symbolic analysis of security protocols,” in Proc. Int. Conf.
Comput. Aided Verification, Springer, 2013, pp. 696–701.
usability. [14] D. Dolev and A. C. Yao, “On the security of public key protocols,” IEEE
Trans. Inf. Theory, vol. 29, no. 2, pp. 198–208, Mar. 1983.
[15] OWASP, “OWASP risk rating methodology,” Aug. 2018. [Online].
Available: https://www.owasp.org/index.php/OWASP_Risk_Rating_
Methodology
A. Future Work [16] European Union, “Directive (EU) 2015/2366,” Official J. Eur. Union,
vol. OJ L 337/35, May 2015. [Online]. Available: http://data.europa.eu/
Given the rising importance of eID cards, we plan to elaborate eli/dir/2015/2366/oj
on how they can be involved in other authentication contexts [17] Ministero dell’Interno, “CIE features,” 2020. [Online]. Available: https:
such as those dealing with OpenID Connect or FIDO2, in order to //www.cartaidentita.interno.gov.it/en/cie/cie-features/
[18] ITU, “Information Technology – Open Systems Interconnection – The
understand the advantages they could bring in terms of security. Directory: Public-Key and Attribute Certificate Frameworks”, Oct. 2006.
We are also going to refine the formal models of our use case [Online]. Available: http://handle.itu.int/11.1002/1000/13031
scenario, possibly moving to more supported model checkers [19] GIXEL, European Card for e-Services and National e-ID Applications,
Feb. 2009.
such as Tamarin [13], which would allow us to benefit from the [20] P. A. Grassi et al., “Digital identity guidelines: Federation and
active community of users. Finally, we would like to improve our assertions,” NIST Special Publication 800–63C. NIST, Jun. 2017,
risk analysis procedure by enhancing its flexibility, for instance doi: 10.6028/NIST.SP.800-63c.
[21] M. Pernpruner, R. Carbone, G. Sciarretta, and S. Ranise, “Complemen-
by adapting the risk factors’ values to the context in order to tary website,” 2022. [Online]. Available: https://st.fbk.eu/complementary/
allow for more granular what-if analyses. TDSC2022
1950 IEEE TRANSACTIONS ON DEPENDABLE AND SECURE COMPUTING, VOL. 21, NO. 4, JULY/AUGUST 2024

[22] G. Sciarretta, R. Carbone, S. Ranise, and L. Viganò, “Formal analysis of Roberto Carbone received the PhD degree in elec-
mobile multi-factor authentication with single sign-on login,” ACM Trans. tronic and computer engineering and telecommuni-
Privacy Secur., vol. 23, no. 3, 2020, Art. no. 13. cations from the University of Genoa, Italy, in 2009.
[23] R. Hajian, S. ZakeriKia, S. Erfani, and M. Mirabi, “SHAPARAK: Scalable He is the head of the Security & Trust Research Unit,
healthcare authentication protocol with attack-resilience and anonymous Center for Cybersecurity, Fondazione Bruno Kessler.
key-agreement,” Comput. Netw., vol. 183, 2020, Art. no. 107567. His previous appointments include a period as visiting
[24] S. S. Sahoo, S. Mohanty, and B. Majhi, “A secure three factor based scholar with the Department of Computer Science,
authentication scheme for health care systems using IoT enabled devices,” University of Pittsburgh, Pennsylvania, US. He has
J. Ambient Intell. Humanized Comput., vol. 12, pp. 1419–1434, 2021. been involved in several international and national
[25] H. Lee, D. Kang, J. Ryu, D. Won, H. Kim, and Y. Lee, “A three-factor research projects and industrial collaborations. His
anonymous user authentication scheme for Internet of Things environ- research focuses on digital identity management and
ments,” J. Inf. Secur. Appl., vol. 52, 2020, Art. no. 102494. the (formal) analysis of security protocols and services.
[26] X. Li, J. Niu, M. Z. A. Bhuiyan, F. Wu, M. Karuppiah, and S. Kumari, “A
robust ECC-based provable secure authentication protocol with privacy
preserving for industrial Internet of Things,” IEEE Trans. Ind. Informat.,
vol. 14, no. 8, pp. 3599–3609, Aug. 2018.
[27] R. Vinoth, L. J. Deborah, P. Vijayakumar, and N. Kumar, “Secure multifac-
tor authenticated key agreement scheme for industrial IoT,” IEEE Internet
of Things J., vol. 8, no. 5, pp. 3801–3811, Mar. 2021.
[28] Y. Guo, Z. Zhang, and Y. Guo, “SecFHome: Secure remote authentication Giada Sciarretta received the MSc degree in math-
in fog-enabled smart home environment,” Comput. Netw., vol. 207, 2022, ematics and the PhD degree in computer science
Art. no. 108818. from the University of Trento, in 2012 and 2018,
[29] Y. Xia, R. Qi, S. Ji, J. Shen, T. Miao, and H. Wang, “PUF-assisted respectively. She is a researcher with the Security
lightweight group authentication and key agreement protocol in smart & Trust Research Unit, Center for Cybersecurity,
home,” Wireless Commun. Mobile Comput., vol. 2022, 2022. Fondazione Bruno Kessler. Her research focuses on
[30] Y. Lu, G. Xu, L. Li, and Y. Yang, “Anonymous three-factor authenticated digital identity with a specialization in the design,
key agreement for wireless sensor networks,” Wireless Netw., vol. 25, security and risk assessment of access delegation and
pp. 1461–1475, 2019. single sign-on protocols (e.g., OAuth 2.0 and OpenID
[31] T.-Y. Wu, L. Yang, Z. Lee, S.-C. Chu, S. Kumari, and S. Kumar, “A Connect), multi-factor authentication (e.g., based on
provably secure three-factor authentication protocol for wireless sensor biometric or eID cards) and fully-remote enrollment
networks,” Wireless Commun. Mobile Comput., vol. 2021, pp. 1–15, 2021. procedures.
[32] Q. Wang and D. Wang, “Understanding failures in security proofs of
multi-factor authentication for mobile devices,” IEEE Trans. Inf. Forensics
Security, vol. 18, pp. 597–612, 2023.
[33] H. Feng, H. Li, X. Pan, and Z. Zhao, “A formal analysis of the FIDO UAF
protocol,” in Proc. Int. Conf. Netw. Distrib. Syst. Secur. Symp., 2021. [On-
line]. Available: https://www.ndss-symposium.org/ndss-paper/a-formal-
analysis-of-the-fido-uaf-protocol/
[34] J. Guan, H. Li, H. Ye, and Z. Zhao, “A formal analysis of the FIDO2 Silvio Ranise received the PhD degree in computer
protocols,” in Proc. Eur. Symp. Res. Comput. Secur., Springer, Switzerland, engineering from the University of Genoa, Italy, and
2022, pp. 3–21. the Henri Poincaré University, Nancy, France. He
[35] D. Fett, R. Küsters, and G. Schmitz, “A comprehensive formal security is the director with the Center for Cybersecurity,
analysis of OAuth 2.0,” in Proc. ACM SIGSAC Conf. Comput. Commun. Fondazione Bruno Kessler and a full professor of
Secur., 2016, pp. 1204–1215. computer science with the University of Trento De-
[36] J. Zhang, L. Yang, W. Cao, and Q. Wang, “Formal analysis of 5G EAP-TLS partment of Mathematics. He has been a researcher
authentication protocol using proverif,” IEEE Access, vol. 8, pp. 23 674–23 with the INRIA-National Institute for Research in
688, 2020. Digital Science and Technology, visiting professor
with the Computer Science Department, University of
Milan and senior researcher with Fondazione Bruno
Kessler.

Marco Pernpruner received the BSc degree in infor-

mation and business organisation engineering from
the University of Trento, in 2016, and the MSc degree
in computer science and engineering from the Univer-
sity of Verona, in 2019. He is a student of the PhD
Program in security, risk and vulnerability, jointly
offered by the University of Genoa and Fondazione
Bruno Kessler (Italy). He has also been a visiting PhD
student with King’s College London (UK), in 2022.
His research focuses on digital identity, with a spe-
cialization in the design, security and risk assessment
of multi-factor authentication and fully-remote enrollment procedures.