Modelling Duqu 2.0 Malware using Attack Trees with Sequential
Conjunction
Peter Maynard, Kieran McLaughlin and Sakir Sezer
Centre for Secure Information Technologies, Queen’s University Belfast, BT3 9DT, Belfast, U.K.
Keywords:
Duqu 2.0, Attack Trees with Sequential Conjunction, SAND, Malware Analysis, Threat Modelling.
Abstract:
In this paper we identify requirements for choosing a threat modelling formalisation for modelling sophisti-
cated malware such as Duqu 2.0. We discuss the gaps in current formalisations and propose the use of Attack
Trees with Sequential Conjunction when it comes to analysing complex attacks. The paper models Duqu
2.0 based on the latest information sourced from formal and informal sources. This paper provides a well
structured model which can be used for future analysis of Duqu 2.0 and related attacks.
1 INTRODUCTION
Threat modelling is used to visualise threats to a sys-
tem or process and provide a method to identify vul-
nerable areas which otherwise might have gone unno-
ticed. Threat modelling can be applied to many differ-
ent circumstances ranging from computer networks,
software life cycle, malware analysis and physical se-
curity. Not only can it provide the analyst with a vi-
sual representation, it is also possible to apply quan-
tification methods to each possible threat path, result-
ing in a ranking of the likelihood that a specific vul-
nerability could be exploited. In most cases threat
modelling will take the form of the abstracted high
level steps an adversary could take when attempting
to compromise a process. This is normally repre-
sented as a graph, a tree or a combination matrix, and,
quantification methods are then applied on top of the
model.
Typically there are two main approaches to threat
modelling, there are formalisms which are derived
from, or extend, the original threat trees, and for-
malisms based on attack graphs. Threat tree based
approaches are designed with the perceived threat at
the top of the diagram, which is referred to as the root
node. Underneath are multiple paths and leaf nodes
which an adversary could take to exploit the preserved
threat. Attack graphs are based on graphs where all
nodes are connected to each other and the values of
the vertices are used to measure the quantitative val-
ues. There is an inherent problem when using a graph
to model attacks, they are typically not able to scale
well.
Duqu 2.0 was identified by Kaspersky Labs
(Kaspersky, 2015) in 2014 and have confirmed that
it is the re-emergence of the Duqu malware which
was found in 2011 which was active for a few months
before it was deactivated apparently by its creators.
Duqu 2.0 was first found to be active about three
weeks before the Iranian nuclear talks were to take
place in the same physical location, with speculation
(Schneier, 2015; Zetter, 2015) on Duqu being used
for espionage by a nation state. Duqu and Duqu 2.0
share large amounts of code with the exception that
Duqu 2.0 was compiled with newer tool chains, it also
has many similarities in the way in which the code is
structured. Duqu 2.0 uses a range of compression and
encryption algorithms to avoid being detected, it also
hijacks existing process’s security clearances to pre-
vent anti-virus software from terminating the process.
Contributions: In this paper we develop a model
of the Duqu 2.0 malware based on the latest analy-
sis of the malware. There is a lot of information re-
garding Duqu 2.0 but no single formal definition has
been published, which can be used to deeply analyse
Duqu’s behaviour in greater detail. Most of the in-
formation regarding Duqu 2.0 is only in news reports
and white papers, which detail only the technical de-
tails of how it works. Modelling Duqu 2.0 using a for-
malised threat model gives future researchers a base
to develop new detection techniques and allow for ap-
plication of quantification methodologies.
The paper is organised as follows: Section 2
explores some implementations of threat modelling
within cyber-physical systems. Section 3 discusses
existing threat modelling formalisations, and we
Maynard, P., McLaughlin, K. and Sezer, S.
Modelling Duqu 2.0 Malware using Attack Trees with Sequential Conjunction.
DOI: 10.5220/0005745704650472
In Proceedings of the 2nd International Conference on Information Systems Security and Privacy (ICISSP 2016), pages 465-472
ISBN: 978-989-758-167-0
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
465
model Duqu 2.0 in section 4 before presenting some
conclusions.
2 BACKGROUND
In this section we discuss the use of threat modelling
in the analysis of recent attacks, and threat modelling
specific to cyber-physical systems.
Bryes et.al (Byres et al., 2004) discuss the use
of attack trees in assessing vulnerabilities in SCADA
systems, the authors develop eleven models each with
there own specific attack goal. The models identify
vulnerabilities within SCADA protocols and in ex-
isting methods of ICS deployment. Although these
vulnerabilities are now well-known, they were not at
the time of publishing. Byres demonstrated that the
use of attack trees is suitable for use within the indus-
try and identify flaws with the formalisation. It also
became evident that threat modelling was in need of
more exploration. Since the paper was published in
2004 there has been an uptake in threat modelling.
Ten et.al. (Ten et al., 2007) again used attack trees
to model threats to SCADA systems and have taken
it a step further than Bryes (Byres et al., 2004). They
have identified an analytical method to measure the
vulnerabilities. The analytical method allows for sys-
tematically evaluating threats and countermeasures.
One of the weaknesses which they identified by using
Attack Trees is there is no way to model the sequence
of the steps, which limits its usefulness. Though the
attack trees approach works well for pentesting and
studying security flaws.
Tanu and Arreymbi (Tanu and Arreymbi, 2010)
model seven types of communication based attacks,
such as replay, DoS, man-in-the-middle and com-
mand injection. Augmented vulnerability tree was
able to model the communication based attacks,
and identify issues with the current SCADA proto-
cols. They recommended the use of message hashing
for authentication and encrypted communication be-
tween all devices. Though the mitigation results are
not novel, they were able to make use of threat mod-
elling within a SCADA environment with a successful
results.
Smith and Ma (Ma and Smith, 2013) have written
about risk assessment in which they model the con-
nections of the network based on the network traffic.
This is done on a rule based system similar to fire-
wall rules, once the network is modelled they iden-
tify what software and services are running on each
node and map vulnerabilities to any CVE and CVSS
found in a CVE database. They are then able to in-
fer the most likely multi attack path using that infor-
mation. This works well for mapping existing sys-
tems and performing a good audit of the network, and
could be a very good first step for a penetration tester.
However, this does not help to identify future or exist-
ing attacks, how they may be performed or how they
can infiltrate the network.
Stuxnet is known as possibly the most advanced
persistent threat ever seen, and it was targeting cyber-
physical systems. Since then there has been more re-
search in the field of SCADA and ICS security, than
in previous years. Duqu/Duqu2.0, Havex, BlackEn-
ergy, and the German steel mill attack are malicious
malware which have been seen to specifically target
critical infrastructure, without analysing these threats
we will not be able to defend against them in the fu-
ture. This is where threat modelling can help identify
features to be used for defending our critical infras-
tructures.
3 EXISTING THREAT
MODELLING APPROACHES
This section details four prominent threat modelling
formalisations designed for specifically modelling at-
tacks such as malware and network intrusions. To-
wards the end of this section we discuss the problems
with each of them.
3.1 Time-dependent Attack Trees
Time dependent attack trees evaluate the probability
of an attack as a function of time. The formalisation
has been formally described in (Arnold et al., 2014).
They make use of the sequential AND operator, as
described in section 3.4 which allows for sequential
operations. Due to the nature of the model, each node
needs to be given specific time values, which is the
expected length of time it would take for that step to
be achieved. It uses the bottom-up algorithm to quan-
tify the results. It is particularly difficult to model an
attack and so the developers have used acyclic phase-
type distribution (APH) expressions to simplify de-
velopment of the models, they provide a prototype
web-based interface for generating the APH expres-
sions. It has been applied to a few test cases, and
appears to be able to handle complex attacks such as
Stuxnet.
3.2 Boolean Logic Driven Markov
Processes
Boolean Logic Driven Markov Processes (BDMP) is
a hybrid formalisation which is based on fault trees
ICISSP 2016 - 2nd International Conference on Information Systems Security and Privacy
466
combined with Markov graphs. Invented for use
within the safety and reliability area (Bouissou and
Bon, 2003) it was later applied to the security industry
by (Pietre-Cambacedes and Bouissou, 2010) in 2010.
Its goal is to find a better trade off between readabil-
ity, modelling power and quantification capabilities
with respect to the existing solutions particularly at-
tack trees. BDMP’s advantages over the traditional
attack trees are its ability to use what they call trig-
gers. Triggers allow modelling of sequences and sim-
ple dependencies by conditionally “activating” sub-
trees of the global structure. Because BDMP is based
on fault trees there is an array of connections avail-
able such as AND, OR and PAND gates, this also
gives the model the advantage that it is easy to un-
derstand and read. In (Kriaa et al., 2012) Kriaa mod-
els Stuxnet using BDMP. Quantification for BDMP is
dependent on how the fault tree is modelled, and this
allows for a very versatile set of metrics. Typical met-
rics include: overall mean-time to success, probabil-
ity of success in a given time, ordered list of attack se-
quences leading to the objective, cost of attacks, han-
dling of boolean indicator and so on. It is also pos-
sible to model defence-centric attributes which reflect
the detection and prevention of the system already in
place, this allows for a more realistic prediction of at-
tack path. There is one tool (KB3-BDMP
1
) capable
of developing BDMP models and performing analy-
sis of the model, development and implementation of
the tool was detailed in (Pietre-Cambacedes and Def-
lesselle, 2011).
3.3 CoPNet
CoPNet is a hybrid threat model which combines at-
tack trees and coloured Petri nets. It was partially de-
fined by (Bouchti and Haqiq, 2012), in this informal
specification their case study is based on a SCADA
network. The case study is a simple SCADA network
with a 3-bus power grid which contains a HMI mon-
itoring the three generators, they model the network
and a range of possible threats then identify the most
likely attack path using their quantification method. It
uses attack trees to model the attack to help in simpli-
fying the development and allow for importing exist-
ing models. Once an attack is modelled using attack
trees, CoPNet has detailed a method which can con-
vert attack trees into coloured Petri nets where they
can then perform the threat analysis. The authors
provide partially working tools available to develop
a CoPNet model. When testing we we’re unable to
generate a usable results from the tools.
1
http://researchers.edf.com/software/kb3-80060.html
3.4 Attack Trees with Sequential
Conjunction
Attack trees with sequential conjunction (SAND) is
an enhancement of attack trees, which were popular-
ized by Schneier (Schneier, 1999). SAND enhances
Attack Trees by defining the use of a sequential AND
operator. This allows for the child nodes to be com-
pleted in sequence adding another level of complexity
without losing the simplicity of attack trees and main-
taining the advantages. SAND was defined in 2015
and has formally been described in (Jhawar et al.,
2015), though this is the most formal definition of
the operator it has been used by other formalisations
previously. There is one primary tool which supports
SAND, called ATSyRA (Pinchinat et al., 2014), a tool
built on top of the Eclipse IDE.
3.5 Problems with Current Threat
Modelling Approaches
To effectively model the Duqu 2.0 malware it was
necessary to identify some requirements upon which
to chose a formalisation. It needs to be easy to un-
derstand in both the raw and visual form. An effec-
tive formalisation must be able to represent sequential
events or dependencies to be able to model a complex
process such as Duqu 2.0. A practical formalisation
also needs to support some form of quantification so
the model can be used for analysis of the malware.
It is also desirable to have a formal specification of
the formalisation to ensure that our model is built to a
correct standard.
Time Dependent attack trees upon first glance ap-
pear to meet all the requirements. They have been
formally defined, support sequential operations and
a working tool is available to help generate models.
Though as the name suggests, the only quantification
which it supports is based on time, and that the model
has to be built using acyclic phase-type distribution
(APH) expressions which abstracts the model, thus
losing information, and reduces the readability of the
model. These two points make it unsuitable for our
application as we wish to develop a model which can
be easily understood and provide a base for further
quantification metrics to be applied.
BDMP has a similar ability to represent sequen-
tial operations by using triggers. It has been formally
defined, though it lacks the readability of traditional
attack trees. BDMP is a hybrid formalisation com-
bining attack trees and Markov graphs, this combina-
tion requires the model designer to have a solid under-
standing of BDMP before they can start working on
a model, as well as the modelling tool being depen-
Modelling Duqu 2.0 Malware using Attack Trees with Sequential Conjunction
467
dant on propitiatory software. Due to the high learn-
ing curve of BDMP over some of the other formalisa-
tions the authors feel that due to BDMP’s complexity
it reduces the likelihood of others continuing to use
and share the produced models.
Visually CoPNet is easy to understand and be-
cause it is based on attack trees, its raw form is also
easy to understand, but there is an extra step which
is needed to convert from attack trees into the model.
The conversion software is currently a prototype. This
extra requirement, along with an informal definition
of the formalisation, does not make it easy for other
analysts to continue working with a model. Further-
more the informal paper (Kriaa et al., 2012) says that
it can handle sequential operations but does not detail
how it can be modelled.
Attack Trees with Sequential AND (SAND) are
based on Attack Trees, which makes comprehending
the raw and visual representations of the formalisation
easier than the other formalisation. It is also a well
known and widely used approach which has been pre-
viously applied to cyber-physical systems. With the
addition of the Sequential AND operator, SAND al-
lows modelling of sequential operations and complex
systems. SAND has been formally defined which al-
lows users to develop a model with reassurance that it
can be expanded in future works.
4 MODELLING DUQU 2.0
Based on the analysis of available threat modelling
techniques in the previous section, we have deter-
mined that SAND should be used to model the Duqu
2.0 malware. In this section we will introduce the ba-
sics of SAND and provide a detailed review of Duqu
2.0.
Attack trees are ordered to allow for systemati-
cally identifying different ways in which a system or
process can be attacked. Attack Trees are ordered
with the attacker’s goal at the top, called the root
node, and subsequent child nodes represent the at-
tacker’s sub-goals. The nodes are connected using
disjunctive (OR), conjunctive (AND) and sequential
conjunctive (SAND). The leaves of the tree represent
the attackers actions.
An example model based on the formalisation
(Jhawar et al., 2015) using SAND is shown in Fig-
ure 1, it details a file server offering ftp, ssh, and rsh
services. The Attack Tree shows the ways which an
attack can get root access. There are two ways ei-
ther without authenticating (no-auth) or by authenti-
cating (auth). The first case (no-auth) the user must
gain privileges then perform a local buffer overflow
Figure 1: Example of the SAND formalisation.
attack (lobf). This is where the sequential AND oper-
ator is used, as the steps must be completed in order
for the attack to succeeded. To gain user privileges,
the attack must first exploit the FTP service so that
they can upload a list of trusted hosts which rsh will
used to allow authentication to the server. In this ex-
ample SAND is executed in left to right order across
the nodes, (note that later this paper it is executed top
to bottom due to the size of the diagrams). The second
method of becoming root is to abuse a buffer over in
both the ssh daemon (ssh) and the RSAREF2 library
(rsa). These nodes are linked with AND, as each of
the steps can happen in any order.
Duqu 2.0 is primarily an information gathering
and exfiltration malware, which has a lot of support
for remaining hidden from detection. It runs com-
pletely within RAM to avoid being detected and it
also leverages anti-virus detection system’s defences
to help remain hidden. It was first found on the net-
work owned by Kaspersky spying on their activity.
It had used three highly complex zero-day exploits.
Not only is it able to hide from detection it has over
one hundred plugins, supporting various functions,
and the ability to encrypt and compress using a wide
range of algorithms. Duqu 2.0 comes in two versions,
full and light, full is around 18MB and contains all
the plugins needed, the light version contains just the
bare minimum with the ability to install plugins. One
of the key features is its ability to create an internal
proxy server for all the infected clients within the net-
work, the traffic is then covertly transferred into and
out of the network without raising suspicion.
We have modelled Duqu using the SAND formal-
isation. The next three sections will discuss the model
in detail, which has been broken down into three man-
ageable parts as follows: Figure 2, Initial Compro-
mise and Lateral Movement, Figure 5, Execution of
the Payload, and Figure 4, Command and Control and
Plugin Operations.
ICISSP 2016 - 2nd International Conference on Information Systems Security and Privacy
468
4.1 Part A - Initial Compromise and
Lateral Movement
It is believed that Duqu was delivered by a targeted
spear-phishing campaign. When a victim received an
email and they downloaded and executed a word doc-
ument resulting in the, then zero-day, exploit CVE-
2014-4148, which is a vulnerability in the windows
TrueType font. The attacker crafts a custom TrueType
font which, when it gets parsed by the operating sys-
tem, allows the attacker to execute arbitrary code with
the security permissions of the kernel, virtually unre-
stricted.
Once the malware has kernel access it attempts
to propagate laterally through the network. Again,
it is able to take advantage of another zero-day ex-
ploit (CVE-2014-6324) to gain domain administrator
privileges using a ’pass the hash’ exploit. Essentially
Figure 2: Part A - Initial Compromise and Lateral Move-
ment.
it requests a token from the Kerberos authentication
server, decrypts it and forges the security groups so
that it looks as if the domain user is in the administra-
tors user group. This gives the malware unrestricted
administrative access to the domain controller and the
rest of the network.
Once Duqu 2.0 has administrator access to the do-
main controller it begins to propagate itself across the
network by remotely installing malicious a MSI file
on domain clients. An MSI file is an executable in-
stallation file used by windows, Duqu 2.0 can create
these executables which contain the necessaries to de-
ploy itself on client machines, Figure 3 shows the MSI
structure. It can install the MSI file remotely with
’msiexec.exe’ or using the task scheduler. There are
two versions of the MSI; one basic in-memory remote
backdoor, and another larger version (18MB) which
contains many of the advanced plugins. The MSI file
is encrypted and compressed using a unique combina-
tion of algorithms, making it hard to apply signature
detection upon the files.
4.2 Part B - Execution of the Payload
This section discusses Figure 5 which is focused on
the payload of the malware.
The MSI’s file structure, as shown in Figure 3,
contains two executable blocks, one is ’Custom ac-
tion dll’ which is used to decrypt and decompress the
payload called ’ActionData0’. ActionData0 also con-
tains encrypted and compressed executable code used
to deploy and manage the malware. Duqu is able to
adapt to its environment by using different payloads,
it can dynamically target CPU architectures as well
as a specific process. Kaspersky was able to identify
five payload containers each with similar configura-
tion. In this paper we have gone into detail on the
variation payload Type L with a focus on the Kasper-
sky anti-virus process as this has been the only pro-
cess covered in such detail. The other payloads are
similar to Type L, we have included the differences in
the model, and briefly discuss them towards the end
of this section.
It first attempts to locate and execute ’api-ms-win-
shell-XXXX.dll’, with ’X’ being a decimal, the num-
ber is based on the machine/boot time. If it can find
that it will generate a PRNG kernel object, this allows
for the value to be unique and easily identified by an-
other process if it uses the same number generator.
If it already exists it will load the value located in it,
and if not it will open a device driver and issue some
IOCTL codes to the driver. This is used by another
module further on.
It starts to search though all the running processes
Modelling Duqu 2.0 Malware using Attack Trees with Sequential Conjunction
469
Figure 3: Kaspersky’s breakdown of Duqu’s MSI structure.
(Kaspersky, 2015).
on the host for a matching string, in our model it
specifically looks for ’avp.exe’. Which is the Kasper-
sky anti-virus process. Once it is located, it will at-
tempt to gain kernel access using CVE-2015-2360,
with the intention of loading the ’KMART.dll’ driver
with kernel level permissions. Loading ’KMART.dll’
with high permissions is done so as to allow hijacking
of the anti-virus’ security tokens.
Now it iterates over a list of hard coded registry
keys, with the intention of finding the location of in-
stalled security software, again in this instance we
are specifically looking for ’avp.exe’. Once the loca-
tion is found, it confirms that the file is executable by
checking the environmental values resolve correctly,
the file will open, and it begins with 0x5A4D (MZ).
With the executable correctly identified it will map
it to memory and patch it to jump back to ’kilf.dll’
where it communicates with ’KMART.dll’, which is
a signed driver with kernel level permissions. Once
all this is done, it now appears to the operating sys-
tem and other processes that, Duqu is actually a pro-
cess owned by ’avp.exe’, the anti-virus. This effec-
tively means that it will be ignored by the anti-virus
software as it believes it is secure, and the defences
which Kaspersky have built into their application will
also be hijacked.
The four alternative payloads, G, I, K, and Q are
very similar to L but with some slight differences.
Figure 4: Part C - Command and Control and Plugin Oper-
ations.
Type G, is almost identical to L, but it skips hijack-
ing the processes’ security tokens and automatically
selects a process with a known configuration. Type I,
is the same a G, but it searches for the processes us-
ing a hashed value. Type K, runs from the context of
the current process and blocks the threat until its com-
plete and Q does the same but runs asynchronously.
4.3 Part C - Command and Control and
Plugin Operations
This section will discuss Figure 4, which is focused
on the command and control and plugins. Command
and control instructions are handled with an array of
possible protocols, HTTP and HTTPs, which covertly
transmits the data as a JPEG/GIF to reduce the likeli-
hood of being detected. Duqu can also use SMB net-
ICISSP 2016 - 2nd International Conference on Information Systems Security and Privacy
470
Figure 5: Part B - Execution of the Payload.
work pipes and generic TCP connections. It forwards
remote desktop connections and controls all aspects
of the malware remotely using these methods.
The plugins are separated into two groups, attack
and reconnaissance. The attack section details some
of Duqu’s offensive plugins which range from re-
motely controlling a desktop, to responding to WPAD
requests, and running a fake SMB server to steal lo-
gin credentials. The reconnaissance describes fea-
tures which can be used to collect information, and is
divided into Network and System features. It is able
to detect when a packet capture tool is being run and
take precautions, it can also enumerate though the net-
work identify machines. It can steal passwords from
web browsers, email clients and the operating system.
It also targets specific file types which are related to
the operation of critical infrastructure.
5 CONCLUSION
In this paper we have modelled the fundamental
mechanisms of the Duqu 2.0 malware in a robust
and formally defined model. The model was built
using the latest information on Duqu 2.0 gathered
from various formal and informal sources. The pa-
per has provided the security community with a com-
plex model, which can be expanded to perform quan-
tification analysis of malware. This paper has iden-
tified features which can potentially be used to de-
tect and prevent such malware from appearing in the
wild. The paper has identified a suitable formalisa-
tion which can be used for future research into mal-
ware analysis as well as network threat modelling. By
using Sequential AND to model Duqu we were able
to generate a model which is easy to work with and
allows us to extract common features for future use.
The advantage of using SAND over existing formali-
Modelling Duqu 2.0 Malware using Attack Trees with Sequential Conjunction
471
sations is it’s simplistic modelling of complex attacks.
The textual representation allows for analysis of the
model without specialist tools or complex syntax.
REFERENCES
Arnold, F., Hermanns, H., Pulungan, R., and Stoelinga, M.
(2014). Time-Dependent Analysis of Attacks. In Prin-
ciples of Security and Trust (POST). Springer.
Bouchti, A. and Haqiq, A. (2012). Modeling cyber-attack
for SCADA systems using CoPNet approach. In
2012 International Conference on Complex Systems
(ICCS), pages 1–6.
Bouissou, M. and Bon, J.-L. (2003). A new formalism that
combines advantages of fault-trees and Markov mod-
els: Boolean logic driven Markov processes. Reliabil-
ity Engineering & System Safety, 82(2):149–163.
Byres, E. J., Franz, M., and Miller, D. (2004). The use of
attack trees in assessing vulnerabilities in scada sys-
tems. In in IEEE Conf. International Infrastructure
Survivability Workshop (IISW 04). Institute for Elec-
trical and Electronics Engineers.
Jhawar, R., Kordy, B., Mauw, S., Radomirovi, S., and
Trujillo-Rasua, R. (2015). Attack Trees with Sequen-
tial Conjunction. In ICT Systems Security and Privacy
Protection, number 455 in IFIP Advances in Informa-
tion and Communication Technology, pages 339–353.
Kaspersky (2015). The Mystery of Duqu 2.0: a sophisti-
cated cyberespionage actor returns.
Kriaa, S., Bouissou, M., and Pietre-Cambacedes, L. (2012).
Modeling the Stuxnet attack with BDMP: Towards
more formal risk assessments. In 2012 7th Interna-
tional Conference on Risk and Security of Internet and
Systems (CRiSIS), pages 1–8.
Ma, Z. and Smith, P. (2013). Determining Risks from Ad-
vanced Multi-step Attacks to Critical Information In-
frastructures. In Critical Information Infrastructures
Security, number 8328, pages 142–154. Springer.
Pietre-Cambacedes, L. and Bouissou, M. (2010). Attack
and Defense Modeling with BDMP. In Kotenko, I.
and Skormin, V., editors, Computer Network Security,
number 6258, pages 86–101. Springer.
Pietre-Cambacedes, L. and Deflesselle, Y. (2011). Security
Modeling with BDMP: From Theory to Implementa-
tion. pages 1 – 8.
Pinchinat, S., Acher, M., and Vojtisek, D. (2014). Towards
Synthesis of Attack Trees for Supporting Computer-
Aided Risk Analysis.
Schneier, B. (1999). Attack Trees. Dr. Dobb’s.
Schneier, B. (2015). Duqu 2.0. Schneier on Security.
Tanu, E. and Arreymbi, J. (2010). An examination of the se-
curity implications of the supervisory control and data
acquisition (SCADA) system in a mobile networked
environment. Proceedings of Advances in Computing
and Technology, (AC&T).
Ten, C.-W., Liu, C.-C., and Govindarasu, M. (2007). Vul-
nerability Assessment of Cybersecurity for SCADA
Systems Using Attack Trees. In IEEE Power Engi-
neering Society General Meeting, 2007, pages 1–8.
Zetter, K. (2015). Kaspersky Finds New Nation-State At-
tackIn Its Own Network.
ICISSP 2016 - 2nd International Conference on Information Systems Security and Privacy
472