A Preliminary Application of Generalized Fault Trees to Security
Daniele Codetta-Raiteri
DiSIT, Computer Science Institute, University of Piemonte Orientale, Viale Michel 11, Alessandria, Italy
Keywords:
Attack Trees, Fault Trees, Petri Nets, Dynamic Gates, Repair Box, Parametric Form, Simulation, Reliability,
Security.
Abstract:
Fault trees (FT) are widespread models in the field of reliability, but they lack of modelling power. So, in the
literature, several extensions have been proposed and introduced specific new modelling primitives. Attack
trees (AT) have gained acceptance in the eld of security. They follow the same notation of standard FT, but
they represent the combinations of actions necessary for the success of an attack to a computing system. In
this paper, we extend the AT formalism by exploiting the new primitives introduced in the FT extensions. This
leads to more accurate models. The approach is applied to a case study: the AT is exploited to represent the
attack mode and compute specific quantitative measures about the system security.
1 INTRODUCTION
Fault Trees (FT) (Sahner et al., 1996) are a
widespread model in the reliability field, and repre-
sent how combinations of component failures (called
basic events) lead to the system failure (top event).
Basic events are Boolean variables whose value turn
from false to true when the component fails. The
intermediate events (subsystem failures) and the top
event are Boolean variables as well, with the same
semantics of basic events, so their value can be deter-
mined by means of Boolean gates (AND, OR, etc.).
From a FT model, we can obtain the minimal cut
sets which are the minimal sets of component fail-
ures (basic events)determining the system failure (top
event). If a probability distribution is associated with
basic events, the FT allows the computation of several
probabilistic measures, such as the system unreliabil-
ity (the probability that the system is failed at a given
time), the probability of minimal cut sets, and impor-
tance (sensitivity) indices. The FT modelling power
is rather limited, mainly because basic events are as-
sumed to be independent. So, in the literature, several
FT extensions have been proposed introducing new
modelling capabilities, as described in Section 2.
Attack trees (AT) (Schneier, 1999) can be consid-
ered the application of FT in the field of security. In
other words, an AT follows the same formalism of a
FT, but the goal is representing the combinations of
actions (basic events) by an attacker, in order to suc-
ceed in compromising a system (top event). AT can
be used to both graphically represent the attack mode,
and assess the system security: both the qualitative
analysis (minimal cut sets detection) and the quantita-
tive analysis (computation of probabilistic measures)
can be performed.
AT typically exploit only Boolean gates in order
to express the attack mode. So, AT and FT have the
same modelling power. In this paper, we propose
to include in an AT model all the modelling primi-
tives proposed in the several FT extensions, with the
goal of designing more accurate FT models express-
ing more complex attack modes. In particular, in Sec-
tion 3, we model and evaluate a case study by means
of an AT including Boolean gates, dynamic gates, re-
pair boxes, and the parametric form. The AT model is
evaluated by means of Petri net (Sahner et al., 1996)
generation and simulation, with the goal of computing
quantitative indices concerning the system security.
The use of simulation instead of analysis is justified
by the presence of repeatable events which lead to an
infinite state space in case of analysis of the model, as
discussed in Section 3.
2 RELATED WORK
An efficient way to perform both the qualitative and
the quantitative analysis (Section 1) of a FT, consists
of the generation and the analysis of the equivalent
Binary Decision Diagram (BDD) (Rauzy, 1993).
One of the ways to improve the reliability of a
609
Codetta-Raiteri D..
A Preliminary Application of Generalized Fault Trees to Security.
DOI: 10.5220/0004612606090614
In Proceedings of the 10th International Conference on Security and Cryptography (SECRYPT-2013), pages 609-614
ISBN: 978-989-8565-73-0
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
system, consists of replicating its critical components
or subsystems; in these cases, the construction and
the analysis of the FT may become quite unpracti-
cal because the model will be composed by several
identical (large) sub-trees representing the replicated
parts. Parametric Fault Trees (PFT) (Bobbio et al.,
2003) were proposed with the purpose of providing
the compact modelling of such parts. Using PFT,
identical sub-trees are folded into a single parametric
sub-tree, while the identity of each replica is main-
tained through the possible values of the parameters.
Dynamic Fault Trees (DFT) (Dugan et al., 1992)
introduce dynamic gates representing several kinds of
dependency among events: functional dependencies,
dependencies concerning the order of events, and the
presence of spare components. Due to the dependen-
cies in the model, DFT need the state space solution;
this means generating all the possible system states
and stochastic transitions between states. In other
words, we need to generate and analyze the Continu-
ous Time Markov Chain (CTMC) (Sahner et al., 1996)
equivalent to the DFT.
Repairable Fault Trees (RFT) (Codetta et al.,
2004) introduce a new primitive called Repair Box
representing the presence of a repair process involv-
ing a certain set of components, and activated by the
occurrence of a specific failure event. This establishes
some dependencies among the failure and the repair
events, so the state space analysis is required, as in
the case of DFT. The state space analysis of a DFT
or RFT can be performed by conversion into a Gen-
eralized Stochastic Petri Net (GSPN) (Sahner et al.,
1996) and by exploiting the available GSPN solution
techniques consisting of the generation and the anal-
ysis of the underlying CTMC. The GSPN model can
undergo simulation as well.
In (Codetta, 2005) the modelling primitives
present in FT, PFT, DFT and RFT formalisms have
been integrated into a single formalism called Gen-
eralized Fault Tree (GFT). So, in a GFT model, we
can exploit in a combined way, the compact mod-
elling of redundancies and symmetries, the dependen-
cies among the events, the repair of components or
subsystems.
There has been a significant amount of work on
developing quantitative evaluation tools for computer
security (MATFIA project, 2003; Dacier et al., 1996a;
Dacier et al., 1996b). In particular, the methodol-
ogy of AT has become popular and has been applied
in several contexts, such as SCADA systems (Byres
et al., 2004; Ten et al., 2007). AT can incorporate
defense mechanisms or countermeasures (Roy et al.,
2012). In (Kordy et al., 2011), the point of view of the
attacker as well as the point of view of the defender
can be analysed.
Besides AT, other modelling formalisms have
been applied to security. In (McDermott, 2000) Petri
nets are exploited to represent the attack mode, while
in (Helmer et al., 2007) an AT is converted into Petri
net with the aim of evaluating the model. Stochastic
Activity Networks (SAN) (Sanders and Meyer, 2001)
are a particular form of Petri nets; they have been ap-
plied to security in (Gupta et al., 2003). Bayesian
networks (Langseth and Portinale, 2007) are applied
in (Frigault et al., 2008; Zhang and Song, 2011; Xie
et al., 2010). Other modelling formalisms are priv-
ilege graphs (Dacier and Deswarte, 1994) and AD-
VISE (ADversary VIew security Evaluation) (LeMay
et al., 2011).
3 THE CASE STUDY
The case study consists of the acquisition by a not au-
thorized user (hacker), of the root password of a Unix
server, by means of a privilege escalation attack. We
assume that the Unix server is periodically character-
ized by two vulnerabilities: v1 is the possibility that a
not authorized entity (hacker) logs in the server; v2 is
the possibility to crack the root password.
The privilege escalation attack is performed in this
way: in the time interval between the occurrence of
v1, and the detection and recovery of v1 by the sys-
tem administrator, one or more hackers may try to
log in the server (event LOGGING
IN). After the de-
tection of v1 (event v1REP), the administrator may
discover and remove the not authorized users logged
in (event DISCOVERING). The undiscovered hackers
keep their presence in the system and may discover
the root password in two ways: 1) trying to crack
the root password (event CRACKING) during the oc-
currence of v2; 2) trying to guess the root password
(event GUESSING); this operation does not require
any vulnerability. Also v2 may be detected and recov-
ered by the system administrator (event v2REP). Both
v1 and v2 may occur again after their recovery. The
server becomes compromised if at least one hacker
succeeds in discovering the root password. We as-
sume that users authorized to log in the server are de-
pendable and never try to discover the root password.
We ignore the actions that an intruder may perform in
the system after the discovery of the root password.
All the events described above may occur if al-
lowed by the system state and after an interval of time
which is a random variable ruled by the negative ex-
ponential distribution. In this case study, the values of
the parameter λ have been chosen in an arbitrary way.
Values closer to reality might be obtained by means of
SECRYPT2013-InternationalConferenceonSecurityandCryptography
610
Table 1: The mean times to occurrence and the corresponding rates for each event in the case study.
Event Description Mean time to occurrence 1/λ (h) Occurrence rate λ (h
1
)
v1 occurrence of v1 24 · 60 0.000694
v2 occurrence of v2 24 · 90 0.000462
v1REP recovery of v1 24 · 10 0.004166
v2REP recovery of v2 24 · 7 0.005952
LOGGING IN attempt to log in 24 · 2 0.020833
CRACKING attempt to crack the root password 24 0.041666
GUESSING attempt to guess the root password 24 · 365 0.000114
DISCOVERING detection and removal of a user logged in 24 0.041666
statistical investigations. Tab. 1 shows the mean time
to occurrence of each event, with the corresponding
parameter λ.
Some events cannot happen before other ones.
For example, the attempt to crack the root password
(eventCRACKING) cannot be performed if the hacker
has not succeeded in logging in and v2 has not oc-
curred. In a similar way, logging in may be attempted
only after the occurrence of v1. These are the tempo-
ral dependencies among the events in this case study
(the symbol specifies that an event must precede
another one):
v1 LOGGING
IN
v1 v1REP
(LOGGING
IN v2) CRACKING
LOGGING IN GUESSING
(LOGGED
IN v1REP) DISCOVERING
v2 v2REP
Some events, once “enabled”, may be repeatable. For
instance, while v1 is occurring, one or more hackers
may log in the server. In a similar way, while v2 is oc-
curring, one ore more hackers may discover the root
password. The occurrence and the recovery of a vul-
nerability are instead alternating events.
3.1 Model Design
Figure 1 shows the AT representing the case study.
This model uses the modelling primitives collected
in the GFT formalism (Section 2). The top event
(TE), the “root” of the AT, represents the situation
where the server is compromised. This happens if at
least one hacker discovers the root password. There-
fore TE is the output of an OR gate connected to
the event ROOT(i), with i = 1, 2,.. .. ROOT(i) rep-
resents the discovery of the root password by the i-
th hacker introduced inside the system. ROOT(i) is
actually a replicator event. This means that the sub-
tree below ROOT(i) is the compact representation of
several sub-trees with the same structure. The iden-
tity of each sub-tree is maintained by the possible
values of the parameter i which is associated with
the events in the sub-tree, with the exception of v1
and v2 which are instead events shared by all the
replicated sub-trees. So, each sub-tree folded in the
parametric sub-tree, concerns the actions by the i-th
hacker. This notation is called parametric form (Sec-
tion 2). ROOT(i) is the output of another OR gate,
so ROOT(i) happens if the i-th intruder succeeds in
cracking (event CRACKED(i)) or guessing the pass-
word (event GUESSED(i)).
In this model, we use three Sequence Enforcing
(SEQ) gates. SEQ is one of the dynamic gates (Sec-
tion 2) and forces its input events to occur in a specific
order. The output event of this gate corresponds to
the last input event in the sequence. The basic event
CRACKING(i) (attempt to crack the password by the
i-th hacker) is connected as second input, to two SEQ
gates. Therefore this event may happen only after the
success of the log-in (event LOGGED
IN(i)) and the
vulnerability v2 (basic event v2). In the same way,
the attempt to log in (event LOGGING
IN(i)) may
happen only after the vulnerability v1 (basic event
v1). Also the event GUESSING(i) (attempt to guess
the password by the i-th hacker) is connected to a
SEQ gate: such event may happen only after the event
LOGGED
IN(i).
In the model, two repair boxes (Section 2) are
present. In an AT, the repair box can be used to model
the recovery of a vulnerability. The time to recov-
ery (repair) is a random variable, so a repair box is
ruled by the negative exponential distribution. In Fig-
ure 1, the repair box called v1REP represents the re-
covery of v1 and the detection of the not authorized
users logged in. For this reason, v1REP is connected
to the event LOGGED
IN(i) (success to log in) due
to the sequence of the basic events v1 (vulnerability
v1) and LOGGING
IN(i) (attempt to log in). The re-
pair box v2REP instead, represents only the recovery
of the vulnerability v2, so it is connected to the basic
event v2. The rates of basic events and repair boxes
are the values of λ in Tab. 1.
3.2 Model Evaluation
In this paper, the AT model in Figure 1 is translated
into a Petri Net, and in particular into the GSPN (Sec-
tion 2) in Figure 2. Both models have been edited by
APreliminaryApplicationofGeneralizedFaultTreestoSecurity
611
AND gate
OR gate
sequence enforcing
gate (SEQ)
Repair Box (RB)
Basic Event (BE)
Intermediate Event (IE)
Replicator Event (RE)
Top Event (TE)
parameter
Types of node
Input/Output arc
Order arc
Types of arc
Figure 1: Attack tree model of the case study, using GFT formalism (the labels of the nodes are explained in Tab. 1).
Figure 2: GSPN model of the case study (the details of the model are reported in (Codetta, 2013)).
means of Draw-Net (Codetta et al., 2006).
In FT, basic events are repeatable only in case of
repair. For instance, a component may fail and then,
undergo repair, fail again, and so on. In an AT, a ba-
sic event may be repeatable for an undefined number
of times, even in absence of recovery. For example,
while the system suffers from the vulnerability v1, an
attempt to log in may occur even if another attempt
SECRYPT2013-InternationalConferenceonSecurityandCryptography
612
has already been done. As a consequence, any num-
ber of hackers may log in. The repetitions of basic
events leads the dimensions of the state space to be-
come infinite, so the model cannot undergo analysis.
A remedy to this problem consists of setting a limit to
the number of repetitions of an event. For instance,
we could assume that 10 is the highest number of
hackers logged in. This approach reduces the dimen-
sions of the state space, but they still remain relevant,
and the model may not be realistic. So, the GSPN
obtained from the AT, has been evaluated using sim-
ulation instead of analysis. In this way, we avoid to
impose limits to the number of event repetitions, and
the simulation execution is less expensive than anal-
ysis, in terms of computing complexity. We executed
100000 simulation cycles in order to obtain the results
described below.
Several measures concerning the system security
as a function of the time have been computed by
means of GSPN simulation. Figure 3 shows the prob-
ability that the system has been compromised. This
means that at least one intruder has discovered the
root password. Figure 4 shows the mean number of
intruders that have discovered the root password, by
means of password cracking or guessing. Other mea-
sures are computed in (Codetta, 2013) where we de-
scribe the details of the GSPN in Figure 2 and the way
to compute the measures on it.
4 CONCLUSIONS
The aim of this paper is transferring the previous ex-
periences about FT extensions, from reliability to se-
curity. The GFT formalism defined for reliability
evaluation purposes, has been adopted for AT, so that
we can model the attack mode by resorting to a sin-
gle generalized formalism including and integrating
Boolean gates, dynamic gates, the parametric form
and repair boxes. In this way, the modelling power of
AT is improved in a relevant way, so that more accu-
rate models can be designed. The approach has been
applied to a case study characterized by dependencies
among events, recoveries and symmetries. The cur-
rent work is a first attempt to use the GFT formalism
for AT, so the case study is rather preliminary; how-
ever, it serves as proof-of-concept to demonstrate the
feasibility of using the GFT formalism, with the con-
sequent improvement of the modelling possibilities.
The approach may be applied to more realistic cases,
for instance by using more accurate probability distri-
bution and rates. Actually the parameters in the case
study have been chosen in an arbitrary way. The AT
of the case study has been evaluated by conversion
Figure 3: Probability that the system is compromised.
Figure 4: The mean number of intruders that have discov-
ered the root password, by means of password cracking,
password guessing, or any of them.
into a Petri net and in particular, a GSPN undergoing
simulation. The goal is to avoid the problem of state
space explosion, due to the repeatable events.
We believe that the formalism needs further im-
provements in order to be suitable for the security
field. For example, using GFT formalism, the AT
model takes into account both the attack mode and
the recovery mode. Actually, repair boxes can repre-
sent reactive recovery processes. This means that the
recovery can be performed only as a consequence of a
partial or complete intrusion. The formalism may be
extended by taking into account the preventive recov-
ery as well. In this way, preventive countermeasures
could be included in the AT model. This was already
done in (Roy et al., 2012; Kordy et al., 2011), but
using only Boolean gates. Moreover, we plan to com-
pute indices which are more security-oriented, with
respect to the measures computed in this paper.
Our intention in the future is solving AT mod-
els by means of Dynamic Bayesian Networks
(DBN) (Portinale et al., 2007), already exploited for
FT analysis. The advantage is the possibility of com-
puting predictive, diagnostic, or importance measures
conditioned by observations about the system or com-
ponents state. In the security field, observations may
concern the action by intruders, the presence of vul-
nerabilities or countermeasures. Importance mea-
APreliminaryApplicationofGeneralizedFaultTreestoSecurity
613
sures for security, based on conditioned probability,
are defined in (Roy et al., 2012). We plan to use AT
as an high-level model to represent the attack mode
and generate the corresponding DBN.
REFERENCES
Bobbio, A., Franceschinis, G., Gaeta, R., and Portinale,
L. (2003). Parametric fault tree for the dependability
analysis of redundant systems and its high-level Petri
net semantics. IEEE Transactions on Software Engi-
neering, 29(3):270–287.
Byres, J., Franz, M., and Miller, D. (2004). The use of
attack trees in assessing vulnerabilities in SCADA
systems. In International Infrastructure Survivability
Workshop, Lisbon.
Codetta, D. (2005). Extended Fault Trees Analysis sup-
ported by Stochastic Petri Nets. PhD thesis, Dipar-
timento di Informatica, Universit`a di Torino.
Codetta, D. (2013). Generalized fault trees: from reliability
to security. Technical report, DiSIT, Istituto di Infor-
matica, Universit`a del Piemonte Orientale.
Codetta, D., Franceschinis, G., and Gribaudo, M. (2006).
Defining formalisms and models in the Draw-Net
Modelling System. In International Workshop on
Modelling of Objects, Components and Agents, pages
123–144, Turku, Finland.
Codetta, D., Franceschinis, G., Iacono, M., and Vittorini,
V. (2004). Repairable Fault Tree for the automatic
evaluation of repair policies. In International Con-
ference on Dependable Systems and Networks, pages
659–668, Florence, Italy. IEEE.
Dacier, M. and Deswarte, Y. (1994). Privilege graph: an ex-
tension to the typed access matrix model. In Computer
Security, pages 319–334. Springer.
Dacier, M., Deswarte, Y., and Kaˆaniche, M. (1996a). Mod-
els and tools for quantitative assessment of operational
security. Information systems security, pages 177–
186.
Dacier, M., Deswarte, Y., and Kaˆaniche, M. (1996b). Quan-
titative assessment of operational security: Models
and tools. Information Systems Security.
Dugan, J. B., Bavuso, S. J., and Boyd, M. A. (1992). Dy-
namic Fault-Tree Models for Fault-Tolerant Computer
Systems. IEEE Transactions on Reliability, 41:363–
377.
Frigault, M., Wang, L., Singhal, A., and Jajodia, S. (2008).
Measuring network security using dynamic bayesian
network. In Workshop on Quality of protection, pages
23–30. ACM.
Gupta, V., Lam, V., Ramasamy, H. V., Sanders, W. H.,
and Singh, S. (2003). Dependability and performance
evaluation of intrusion-tolerant server architectures.
In Dependable Computing, pages 81–101. Springer.
Helmer, G., Wong, J., Slagell, M., Honavar, V., Miller, L.,
Wang, Y., Wang, X., and Stakhanova, N. (2007). Soft-
ware fault tree and coloured petri net–based specifica-
tion, design and implementation of agent-based intru-
sion detection systems. International Journal of In-
formation and Computer Security, 1(1):109–142.
Kordy, B., Mauw, S., Radomirovi´c, S., and Schweitzer, P.
(2011). Foundations of attack–defense trees. Formal
Aspects of Security and Trust, pages 80–95.
Langseth, H. and Portinale, L. (2007). Bayesian Net-
works in reliability. Reliability Engineering and Sys-
tem Safety, 92:92–108.
LeMay, E., Ford, M. D., Keefe, K., Sanders, W. H., and
Muehrcke, C. (2011). Model-based security metrics
using adversary view security evaluation (advise). In
International Conference on Quantitative Evaluation
of Systems, pages 191–200. IEEE.
MATFIA project (2000-2003). Malicious-and
accidental-fault tolerance for internet applica-
tions. http://research.cs.ncl.ac.uk/cabernet/
www.laas.research.ec.org/maftia/.
McDermott, J. P. (2000). Attack Net Penetration Testing. In
Workshop on New security paradigms.
Portinale, L., Bobbio, A., Codetta-Raiteri, D., and Mon-
tani, S. (2007). Compiling dynamic fault trees into
dynamic Bayesian nets for reliability analysis: The
Radyban tool. CEUR Workshop Proceedings, 268.
Rauzy, A. (1993). New Algorithms for Fault Trees
Analysis. Reliability Engineering & System Safety,
05(59):203–211.
Roy, A., Kim, D. S., and Trivedi, K. S. (2012). Attack coun-
termeasure trees (ACT): towards unifying the con-
structs of attack and defense trees. Security and Com-
munication Networks.
Sahner, R., Trivedi, K., and Puliafito, A. (1996). Perfor-
mance and Reliability Analysis of Computer Systems;
An Example-based Approach Using the SHARPE Soft-
ware Package. Kluwer Academic Publisher.
Sanders, W. and Meyer, J. (2001). Stochastic activity
networks: Formal definitions and concepts. Lecture
Notes in Computer Science, 2090:315–343.
Schneier, B. (1999). Attack trees. Dr. Dobb Journal of
Software Tools, 24(12):21–29.
Ten, P. C.-W., Liu, C.-C., and Govindarasu, M. (2007).
Vulnerability assessment of cybersecurity for SCADA
systems using attack trees. In Power Engineering So-
ciety General Meeting, pages 1–8.
Xie, P., Li, J. H., Ou, X., Liu, P., and Levy, R. (2010). Us-
ing bayesian networks for cyber security analysis. In
International Conference on Dependable Systems and
Networks, pages 211–220. IEEE.
Zhang, S. and Song, S. (2011). A novel attack graph poste-
rior inference model based on bayesian network. Jour-
nal of Information Security, 2(1):8–27.
SECRYPT2013-InternationalConferenceonSecurityandCryptography
614