Keeping Intruders at Large
A Graph-theoretic Approach to Reducing the Probability of Successful Network
Intrusions
Paulo Shakarian
1
, Damon Paulo
2
, Massimiliano Albanese
3
and Sushil Jajodia
3,4
1
Arizona State University, Tempe, AZ, U.S.A.
2
U.S. Military Academy, West Point, NY, U.S.A.
3
George Mason University, Fairfax, VA, U.S.A.
4
The MITRE Corporation, McLean, VA, U.S.A.
Keywords:
Moving Target Defense, Adversarial Modeling, Graph Theory.
Abstract:
It is well known that not all intrusions can be prevented and additional lines of defense are needed to deal with
intruders. However, most current approaches use honeynets relying on the assumption that simply attracting
intruders into honeypots would thwart the attack. In this paper, we propose a different and more realistic
approach, which aims at delaying intrusions, so as to control the probability that an intruder will reach a
certain goal within a specified amount of time. Our method relies on analyzing a graphical representation
of the computer network’s logical layout and an associated probabilistic model of the adversary’s behavior.
We then artificially modify this representation by adding “distraction clusters” collections of interconnected
virtual machines at key points of the network in order to increase complexity for the intruders and delay the
intrusion. We study this problem formally, showing it to be NP-hard and then provide an approximation algo-
rithm that exhibits several useful properties. Finally, we present experimental results obtained on a prototypal
implementation of the proposed framework.
1 INTRODUCTION
Despite significant progress in the area of intrusion
prevention, it is well known that not all intrusions
can be prevented, and additional lines of defense are
needed in order to cope with attackers capable of
circumventing existing intrusion prevention systems.
However, most current approaches are based on the
use of honeypots, honeynets, and honey tokens to
lure the attacker into subsystems containing only fake
data and bogus applications. Unfortunately, these
approaches rely on the unrealistic assumption that
simply attracting an intruder into a honeypot would
thwart the attack. In this paper, we propose a totally
different and more realistic approach, which aims at
delaying an intrusion, rather than trying to stop it, so
as to control the probability that an intruder will reach
This work was partially supported by the Army Re-
search Office under award number W911NF-13-1-0421.
Paulo Shakarian and Damon Paulo were supported by the
Army Research Office project 2GDATXR042. The work of
Sushil Jajodia was also supported by the MITRE Sponsored
Research Program.
a certain goal within a specified amount of time and
keep such probability below a given threshold.
Our approach is aligned with recent trends in cy-
ber defense research, which has seen a growing in-
terest in techniques aimed at continuously changing a
system’s attack surface in order to prevent or thwart
attacks. This approach to cyber defense is generally
referred to as Moving Target Defense (MTD) (Jajodia
et al., 2013) and encompasses techniques designed to
change one or more properties of a system in order
to present attackers with a varying attack surface
2
, so
that, by the time the attacker gains enough informa-
tion about the system for planning an attack, its attack
surface will be different enough to disrupt it.
In order to achieve our goal, our method relies on
analyzing a graphical representation of the computer
network’s logical layout and an associated probabilis-
tic model of the adversary’s behavior. In our model,
an adversary can penetrate a system by sequentially
gaining privileges on multiple system resources. We
2
Generally, the attack surface refers to system resources
that can be potentially used for an attack.
19
Shakarian P., Paulo D., Albanese M. and Jajodia S..
Keeping Intruders at Large - A Graph-theoretic Approach to Reducing the Probability of Successful Network Intrusions.
DOI: 10.5220/0005013800190030
In Proceedings of the 11th International Conference on Security and Cryptography (SECRYPT-2014), pages 19-30
ISBN: 978-989-758-045-1
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
model the adversary as having a particular target (e.g.,
an intellectual property repository) and show how to
calculate the probability of him reaching the target in
a certain amount of time (we also discuss how our
framework can be easily generalized for multiple tar-
gets). We then modify our graphical representation
by adding “distraction clusters” collections of inter-
connected virtual machines at key points of the net-
work in order to reduce the probability of an intruder
reaching the target. We study this problem formally,
showing it to be NP-hard and then provide an approx-
imation algorithm that possesses several useful prop-
erties. We also describe a prototypal implementation
and present our experimental results.
Related Work. Moving Target Defense (MTD) (Ja-
jodia et al., 2013; Jajodia et al., 2011; Evans et al.,
2011) is motivated by the asymmetric costs borne by
cyber defenders. Unlike prior efforts in cyber secu-
rity, MTD does not attempt to build flawless systems.
Instead, it defines mechanisms and strategies to in-
crease complexity and costs for attackers. The recent
trend of high-profile cyber-incidents resulting in sig-
nificant intellectual property theft (Shakarian et al.,
2013) indicates that current practical approaches may
be insufficient.
MTD differs from current practical approaches
which primarily rely on three aspects: (a) attempting
to remove vulnerabilities from software at the source,
(b) patching software as rapidly as possible, and (c)
identifying attack code and infections. The first ap-
proach is necessary but insufficient because of the
complexity of software. The second approach is stan-
dard practice in large enterprises, but has proven dif-
ficult to keep ahead of the threat, nor does it provide
protection against zero-day attacks. The last approach
is predicated on having a signature of malicious at-
tacks, which is not always possible.
MTD approaches aiming at selectively altering a
system’s attack surface (Manadhata and Wing, 2011)
are relatively new. In Chapter 8 of (Jajodia et al.,
2011), Huang and Ghosh present an approach based
on diverse virtual servers, each configured with a
unique software mix, producing diversified attack sur-
faces. In Chapter 9, Al-Shaer investigates an ap-
proach to enables end-hosts and network devices to
change their configuration (e.g., IP addresses). In
Chapter 6, Rinard describes mechanisms to change a
system’s functionality in ways that eliminate security
vulnerabilities while leaving the system able to pro-
vide acceptable functionality. A game-theoretic ap-
proach to increase complexity for the attacker is pre-
sented in (Sweeney and Cybenko, 2012).
The efforts that are more closely related to our
work are those based on the use of honeypots. How-
ever, such approaches significantly differ from our
work in that they aim at either capturing the at-
tacker and stopping the attack (Abbasi et al., 2012)
or collecting information about the attacker for foren-
sic purposes (Chen et al., 2013). There is also a
relatively new corpus of work on attacker-defender
models for cyber-security using game-theoretic tech-
niques: an overview is provided in (Alpcan and Baar,
2010). Work in this area related to this paper include
(Williamson et al., 2012) and (P
´
ıbil et al., 2012). The
work presented in (Williamson et al., 2012) is similar
to our work in that it models the adversary as moving
through a graphical structure. However, that work dif-
fers in that the defender is trying to learn about the at-
tacker’s actions for forensic analysis purposes. In this
work, we do not assume a forensic environment and
rather than trying to understand the adversary, we are
looking to delay him from obtaining access to certain
machines (e.g., intellectual property repositories). In
(P
´
ıbil et al., 2012), the authors use game theoretic
techniques to create honeypots that are more likely
to deceive (and hence attract) an adversary. We view
their approach as complementary to ours, specifically
with regard to the creation of distraction clusters.
2 TECHNICAL PRELIMINARIES
In this section, we first introduce the notion of in-
truder’s penetration network, and then provide a for-
mal statement of the problem we address in the pa-
per. Note that we model a complex system as a set
S = {s
1
,...,s
n
} of computer systems. Each system
in S is associated with a level of access obtained by
the intruder denoted by a natural number in the range
L = {0, . . . ,
max
}. The level of access to a given sys-
tem changes over time, which is treated as discrete
intervals in the range 0,...,t
max
.
For a given system s and level , we shall use a
system-level pair (s,) to denote that the intruder cur-
rently has level of access on system s. We shall use
S to denote the set of all system-level pairs.
Definition 1
(Intruder’s penetration network (
IPN
))
.
Given a system S = {s
1
,...,s
n
}, the intruder’s pen-
etration network for S is a directed graph IPN =
(S,R,π,f), where S is the set of nodes representing
individual computer systems, R S × S is a set of di-
rected edges representing relationships among those
systems. For a given s
i
S, η
i
= {(s
i
,s
) R}. We
define the conditional success probability function
π : S ×S [0,1] as a function that, given two system-
level pairs (s,),(s
,
) returns the probability that an
intruder with access level on s will gain access level
on s
in the next time step (provided that the attacker
SECRYPT2014-InternationalConferenceonSecurityandCryptography
20
Attack Source Inside Router Inside Switch
IA Switch #1 Oracle: DBSRV2
Figure 1: Sample network based on a real-world case: an
attacker targeting the Oracle server penetrates the network
exploiting a vulnerability in the Router.
selects s
as the next target). This function must have
the following properties:
((s,s
)R)(ℓ>0)(π((s,0),(s
,ℓ))=0) (1)
((s,s
)/R)(ℓ,ℓ
>0)(π((s,ℓ),(s
,ℓ
))=0) (2)
(
k
L )(
i
j
π((s,ℓ
i
),(s
,ℓ
k
))π((s,ℓ
j
),(s
,ℓ
k
))) (3)
We define f : S×S as a function that provides
the “fitness” of a relationship. The intuition behind
the fitness f((s
i
,
i
),(s
j
,
j
)) is that it is associated with
the desirability for the attacker (who is currently on
system s
i
with level
i
) to achieve level of access
j
on
system s
j
. If (s
i
,s
j
) / R then f((s
i
,
i
),(s
j
,
j
)) = 0.
We note that, there are mature pieces of software
for generating the graphical structure of the IPN along
with the success probability and fitness function such
as Cauldron and Lincoln Labs’ NetSpa. Additionally,
there are vulnerability databases that can aide in the
creation of an IPN as well. For instance, in NIST’s
NVD database
3
, impact and attack difficulty can map
to fitness and the inverse of the probability of suc-
cess. While we are currently working with Cauldron
to generate the IPN, we do not focus on the creation of
the structure in this paper, but rather on reducing the
overall probability of success of the intruder.
We assume that if a user has no access (i.e., level 0
access) to a system s then the probability of success-
fully infiltrating another system s
from that system
is 0, which is why property 1 in Definition 1 above
is valid and necessary. Similarly, property 2 articu-
lates the fact that if no edge exists between s and s
then the likelihood of a successful attack on s
origi-
nating from s is also 0. Finally, property 3 defines the
intuitive assumption that if an attacker can complete
an attack with a certain probability of success then
if he conducts the same attack with a higher level of
permissions on the original system his probability of
success must be at least the same as it was in the orig-
inal attack.
Example 2.1. Consider the simple network displayed
in Fig. 1 which represents a subset of a real world
3
http://nvd.nist.gov/
network we examined. In this network a user can
have one of two levels of access on each system; he
can have guest privileges (
1
= 1) or root privileges
(
2
= 2). The attacker begins with root privileges
on his personal device (s
1
,2). The network is dis-
played in full in Fig. 2 where nodes represent system-
level pairs and edges represent logical connections
between them. All transitions between system-level
pairs will either involve transitioning to a new system
or to a higher level on the current system. Edges rep-
resenting gaining higher access on the current sys-
tem are red and bold in Fig. 2. Given two system-
level pairs ((s
i
,
i
),(s
j
,
j
)), the fitness of the relation-
ship
4
between them (f((s
i
,
i
),(s
j
,
j
))) is shown on
the edge between them as f, and the conditional suc-
cess probability function (π((s
i
,
i
),(s
j
,
j
))) is shown
as π. For example, in the sample network, when the
attacker begins with root access on his personal com-
puter (s
1
,2) he can gain guest privileges on the inside
router (s
2
,1) with a fitness of 1 and a probability of
success of 0.8. For ease of reading, for all system-
level pairs where f((s
i
,
i
),(s
j
,
j
)) = 0 the edge is not
displayed. The probability of a successful attack oc-
curring is the product of the probability that the at-
tack succeeds and the probability that the attack is
selected by the intruder. In our sample network, then,
when the intruder has guest privileges on the inside
router (s
2
,1) his probability of successfully gaining
root access on that router (s
2
,2) in the next time step
is
1
1+1
× 0.6 = 0.3.
Penetration Sequence. A penetration se-
quence is simply a sequence of system-level
pairs (s
0
,
0
),...,(s
n
,
n
) such that for each
a = (s
i
,
i
),b = (s
i+1
,
i+1
) in the sequence we have
r = (s
i
,s
i+1
) R , π(a,b), f (a, b) > 0. For a sequence
σ we shall denote the number of system-level
pairs with the notation |σ|. For a given sequence
σ, let σ
m
be the sub-sequence of σ consisting of
the first m 1 system-level pairs in σ. For a se-
quence σ = (s
0
,
0
),...,(s
n
,
n
), curSys(σ) = s
n
,
curLvl(σ) =
n
and cur(σ) = (s
n
,
n
). We use the
notation next(σ) to denote the set of system-level
pairs that could occur next in the sequence. Formally:
next(σ)={(s,ℓ)S | ℓ>0(curSys(σ),s)R
̸∃
s.t. (s,ℓ
)σ} (4)
4
We note that in this particular example we have set the
fitness values to all be one. Note that this is just one way
to specify a fitness function - perhaps in the case with no
information about desirability of a system (i.e. akin to using
uniform priors). All of our results and algorithms are much
more general and allow for arbitrary fitness values.
KeepingIntrudersatLarge-AGraph-theoreticApproachtoReducingtheProbabilityofSuccessfulNetworkIntrusions
21
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
2
,2
Inside
Router
S
3
,2
Inside
Switch
f = 1, π = 0.8
S
4
,1
IA Switch
#1
S
4
,2
IA Switch
#1
f = 1, π = 0.8
f = 1, π = 0.7
f = 1, π = 0.6
f = 1, π = 0.3
f =1, π = 0.8
S
5
,2
Oracle:
DBSRV2
*
S
3
,1
Inside
Switch
S
5
,1
Oracle:
DBSRV2
f = 1, π = 0.7
f = 1, π = 0.7
Figure 2: The sample network. Nodes are system-level pairs.
Example 2.2. Fig. 3 depicts all five possible pene-
tration sequences by which the attacker can gain root
access to the Oracle: DBSRV2 (s
5
,2) in our sample
network in five time steps or fewer. The penetration
sequences are labeled as σ
1
through σ
5
.
Model of the Intruder’s Actions. Now, we shall de-
scribe our model. Consider an attacker who has infil-
trated through a sequence of systems specified by σ.
If the penetration is to continue, the intruder must se-
lect a system-level pair from next(σ). The intruder
selects exactly one system-level pair (s, ) with the
following probability:
f(cur(σ), (s, ))
(s
,ℓ
)next(σ)
f(cur(σ), (s
,
))
(5)
Hence, the probability of selection is proportional
to the relative fitness of (s, ) compared to the other
options for the attacker. This aligns with our idea of
fitness: an intruder will attempt to gain access to sys-
tems that are more “fit” with respect to his expertise,
available tools, desirability of the next system, etc.
Note that this probability of selection is not tied to
the intruder’s probability of success. In fact, we con-
sider the two as independent. Hence, the probability
that an intruder selects and successfully reaches (s,)
can be expressed as follows:
f(cur(σ), (s, ))π(cur(σ),(s,))
(s
,ℓ
)next(σ)
f(cur(σ), (s
,
))
(6)
Hence, given that the attacker starts at a certain
(s,), we can compute the sequence probability or
probability of taking sequence σ (provided that σ
starts at (s, ) this probability would be zero oth-
erwise).
|σ|−2
i=0
f(cur(σ
i
),cur(σ
i+1
))π(cur(σ
i
),cur(σ
i+1
))
(s,ℓ)next(σ
i
)
f(cur(σ
i
),(s,))
(7)
Hence, for a given initial (s,) and ending (s
,
),
and length t + 1, we can compute the probability of
starting at (s,) and ending at (s
,
) in t time-steps or
less by taking the sum of the sequence probabilities
for all valid sequences that meet that criterion. For-
mally, we shall refer to this as the penetration prob-
ability and for a given IPN,t,(s,),(s
,
) we shall de-
note this probability as Pen
t
IPN
((s,ℓ),(s
,ℓ
)). Intuitively,
Pen
t
IPN
((s,ℓ),(s
,ℓ
)) is the probability that an attacker at
system s with level of access reaches system s
with
level of access
or greater in t time steps or less.
Example 2.3. The probability of the attacker suc-
cessfully gaining access to (s
5
,2) in five time steps
or fewer is the sum of the probabilities of each of
the five possible penetration sequences depicted in
Fig. 3. For each penetration sequence σ
n
the prob-
ability of the attacker successfully gaining access to
(s
5
,2) through that particular sequence is p
n
. For the
sample network: p
1
= 0.023, p
2
= 0.021, p
3
= 0.021,
p
4
= 0.004, p
5
= 0.016. Thus, the total probability of
a successful attack occurring on system 5 at level 2 in
three time steps or fewer is Pen
t
IPN
(s
1
,2) = 0.085.
3 DISTRACTION CHAINS AND
CLUSTERS
We now introduce the idea of a distraction chain. A
distraction chain is simply a sequence of decoy sys-
tems that we wish to entice an adversary to explore to
distract him from the real systems of the network. In
order to entice the adversary to explore a distraction
chain, we propose adding one-way distraction clus-
ters to S. Hence, the adversary enters such a distrac-
tion cluster and is delayed from returning to the ac-
tual network for a number of time steps proportional
to the size of the cluster. Ideally, a distraction clus-
ter would be large enough to delay the attacker for
a long time; however, larger distraction clusters will
obviously require more resources to construct. Dis-
traction clusters differ from honeypots because they
do not prevent intruders from reaching other portions
of the network, thus minimizing the risk of the in-
truder realizing that he is trapped. Again, the goal
of a honeypot is to prevent an attacker from complet-
ing his attack by trapping him under the assumption
SECRYPT2014-InternationalConferenceonSecurityandCryptography
22
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
4
,2
IA Switch
#1
S
5
,2
Oracle:
DBSRV2
S
3
,1
Inside
Switch
S
5
,1
Oracle:
DBSRV2
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
4
,2
IA Switch
#1
S
5
,2
Oracle:
DBSRV2
S
3
,1
Inside
Switch
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
4
,2
IA Switch
#1
S
5
,2
Oracle:
DBSRV2
S
3
,1
Inside
Switch
S
4
,1
IA Switch
#1
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
4
,2
IA Switch
#1
S
5
,2
Oracle:
DBSRV2
S
3
,1
Inside
Switch
S
3
,2
Inside
Switch
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
4
,2
IA Switch
#1
S
5
,2
Oracle:
DBSRV2
S
3
,1
Inside
Switch
S
2
,2
Inside
Router
0.8
0.8
0.8
0.8
0.8
0.8
0.7
0.35 0.233 0.35
0.233 0.35
0.35
0.35
0.233
0.45
0.267
0.1 0.4
0.35
0.35
0.35 0.8 0.3
p
1
= 0.023
p
2
= 0.021
p
3
= 0.021
p
4
= 0.004
p
5
= 0.016
σ
1
σ
2
σ
3
σ
4
σ
5
Figure 3: All possible penetration sequences with ve time steps or fewer.
that once he is trapped he will not leave it, while the
goal of a distraction cluster is more realistically
to delay the attacker in order to reduce the probabil-
ity that he will successfully complete his attack in a
given amount of time.
Clearly, a valid distraction cluster would be con-
nected to the rest of the system through a one-way
connection and the cluster must be created in a man-
ner where there is at least one (preferably many) dis-
traction chain of the necessary length (based on an ex-
pected limit of time we expect the intruder to remain
in the network before discovery).
For now we shall leave the creation of distraction
clusters to future work (e.g., the work of (P
´
ıbil et al.,
2012) may provide some initial insight into this prob-
lem) and instead focus on the problem of adding dis-
traction clusters to a system. The configuration of the
first system of a distraction cluster will be a key ele-
ment in setting up a distraction chain. In particular,
the open ports, patch level, installed software, operat-
ing system version, and other vulnerabilities present
on that lead system, as well as any references of that
system found elsewhere on the network will dictate
how “fit” an attacker will determine such a system
to be and the probability of success he will have in
entering into the distraction chain. Note that the fit-
ness of this first system cannot be arbitrarily high and
should be considered based on a realistic assessment
of why the attacker would select such a system. Fur-
ther, the probability of an adversary obtaining privi-
leges on such a system should be set in such a way
where it is not overly simple for the intruder to gain
access - or he might suspect it is a decoy. Addition-
ally, the last system in the distraction cluster must be
configured in a way to reconnect it to the actual net-
work.
Throughout the paper, we will consider a set of
configurations available to the defender denoted CFG.
For instance, CFG may consist of a predetermined set
of virtual machine images available to the security
team. In addition we will consider a set of potential
distraction clusters denoted CL. For each cl CL
there exists value t
cl
N, a natural number equal to
the minimum number of time steps elapsed before
an attacker is able to leave the cluster and return
to the network. For each cfg CFG, for the lead
and last systems in the distraction cluster (resp.
s
dc1
, s
dc2
) there are associated conditional prob-
abilities (resp. π
cfg,cl
: S × {(s
dc1
,)} [0,1],
π
cfg,cl
: {(s
dc2
,)} × S [0,1]) and fitness
function (resp. f
cfg,cl
: S × {(s
dc1
,)} ,
f
cfg,cl
: {(s
dc2
,)} × S ). These functions
are based on the software installed on and the vul-
nerabilities present in that particular configuration.
Hence, once a distraction cluster is added it contains
a lead system configured with configuration cfg.
The resulting IPN formed with the addition of
distraction cluster includes conditional probability
and fitness functions that are the concatenation of
π,π
cfg,cl
and f,f
cfg,cl
respectively. Additionally, for
each s S we add (s,s
dc1
) to R where there exists
s, where > 0 s.t. π
cfg,cl
((s,),(s
dc1
,1)) > 0
and f
cfg,cl
((s,),(s
dc1
,1)) > 0 and we add
(s
dc2
,s) to R where there exists s, where
> 0 s.t. π
cfg,cl
((s
dc2
,1),(s,)) > 0 and
f
cfg,cl
((s
dc2
,1),(s,)) > 0. In other words, a
logical connection is formed from all systems in
S for which, if connected to s
dc1
or s
dc2
there is a
non-zero probability that the intruder can gain a level
of access greater than = 0. We can easily restrict
which relationship are added by modifying the f
cfg,cl
functions. For a given IPN, set of distraction clusters,
and set of configuration-cluster pairs PCP CFG × CL,
we will use the notation IPN PCP to denote the
concatenation of the intrusion penetration network
and the set of configuration-cluster pairs.
KeepingIntrudersatLarge-AGraph-theoreticApproachtoReducingtheProbabilityofSuccessfulNetworkIntrusions
23
Adding Distraction Clusters. We now have the
pieces we need to introduce the formal definition of
our problem.
Definition 2 (Cluster Addition Problem). Given IPN
(S,R,π,f), systems s,s
S, access levels
s
,
s
L ,
set of potential distraction clusters CL, set of configu-
rations CFG, natural number k, real number x [0,1],
and time-limit t, find PCP CFG × CL s.t. |PCP| k and
Pen
t
IPN
((s,ℓ),(s
,ℓ
)) Pen
t
IPNPCP
((s,ℓ),(s
,ℓ
)) > x.
Example 3.1. Following along with our sample net-
work, the set of configurations, CFG, displayed in
Fig. 4, and with k = 2, t = 5, x = 0.05, and CL = {cl}
(with t
cl
= 6) we find that PCP = {(cfg
1
,cl),(cfg
3
,cl)}
is a solution to the Chain Addition Problem because
Pen
t
IPN
(s
1
,2) Pen
t
IPNPCP
(s
1
,2) = 0.063 > 0.05 = x
The modified IPN is displayed in Fig. 5.
Unfortunately, this problem is difficult to solve ex-
actly by the following result (the proof is provided in
the appendix).
Theorem 1. The Cluster Addition Problem is NP-
hard and the associated decision problem is NP-
Complete when the number of sequences from (s,)
to (s
,
) is a polynomial in the number of nodes in
the intruder penetration network.
For a given instance of the Cluster Addition
Problem, for a given PCP CFG × CL let orc(PCP) =
Pen
t
IPN
((s,ℓ),(s
,ℓ
)) Pen
t
IPNPCP
((s,ℓ),(s
,ℓ
)). In the opti-
mization version of this problem, this is the quan-
tity we attempt to optimize. Unfortunately, as a by-
product of Theorem 1 and the results of (Feige, 1998)
(Theorem 5.3), there are limits to the approximation
we can be guaranteed to find in polynomial time.
Theorem 2. With a cardinality constraint, finding set
PCP s.t. orc(PCP) cannot be approximated in PTIME
within a ratio of
e1
e
+ ε for some ε > 0 (where e is
the base of the natural log) unless P=NP.
However, orc does have some useful properties.
Lemma 1 (Monotonicity). For PCP
PCP CFG ×CL,
orc(PCP
) orc(PCP).
Lemma 2 (Submodularity). For PCP
PCP CFG ×
CL and pc = (cfg, cl) / PCP we have:
orc(PCP {pc})orc(PCP) orc(PCP
{pc})orc(PCP
)
4 ALGORITHMS
Greedy Approach. Now we introduce our greedy
heuristic for the Chain Addition Problem.
Algorithm 1: GREEDY CLUSTER.
Require: Systems s,s
, access levels
s
,
s
, set of distrac-
tion chains CL, set of protocols CFG, natural number k
and time limit t
Ensure: Subset PCP CFG × CL
1: PCP =
/
0
2: while |PCP| k do
3: curBest = null, curBestScore = 0
4: for (cfg, cl) (CFG × CL) PCP do
5: curScore = orc(PCP {(cfg,cl)}) orc(PCP)
6: if curScore curBestScore then
7: curBest = (cfg,cl)
8: curBestScore = curScore
9: end if
10: end for
11: PCP = PCP {curBest}
12: end while
13: return PCP.
Example 4.1. When run on our sample net-
work, GREEDY CLUSTER selects (cfg
1
,cl) in the
first iteration of the loop at line 4, lowering
Pen
t
IPNPCP
(s
1
,2) from 0.085 to 0.037. In the next
iteration GREEDY CLUSTER selects (cfg
3
,cl), low-
ering Pen
t
IPNPCP
(s
1
,2) from 0.037 to 0.022. At this
point, since |PCP| = 2 = k, GREEDY CLUSTER re-
turns PCP = {(cfg
1
,cl),(cfg
3
,cl)}, a solution to the
Chain Addition Problem under our given constraints.
Though simple, GREEDY CLUSTER, can provide
the best approximation guarantee unless P=NP under
the condition that orc can be solved in PTIME. Con-
sider the following theorem.
Theorem 3. GREEDY CLUSTER provides the best
PTIME approximation of orc unless P=NP if orc can
be solved in PTIME.
However, the condition on solving orc in PTIME
may be difficult to obtain in the case of a very gen-
eral IPN. Though we leave the exact computation of
this function as an open problem, we note that the
straight-forward approach for computation would im-
ply the need to enumerate all sequences from (s,)
to (s
,
) - which can equal the number of t-sized (or
smaller) permutations of the elements of S - a quan-
tity that is not polynomial in the size of IPN. Hence,
a method to approximate orc is required in practice.
Our intuition is that the expensive computation of the
penetration probability can be approximated by sum-
ming up the sequence probabilities of the most prob-
able sequences from (s, ) to (s
,
). The intuition of
approximating the probability of a path between two
nodes in a network (given a diffusion model) based on
high-probability path computation was introduced in
(Chen et al., 2010) where it was applied to the maxi-
mum influence problem. However, our approach dif-
fers in that (Chen et al., 2010) only considered the
SECRYPT2014-InternationalConferenceonSecurityandCryptography
24
Figure 4: The set CFG of possible configurations available to the defender.
S
1
,2
Attack
Origin
S
2
,1
Inside
Router
S
2
,2
Inside
Router
S
3
,2
Inside
Switch
f = 1, π = 0.8
S
4
,1
IA Switch
#1
S
4
,2
IA Switch
#1
f = 1, π = 0.8
f = 1, π = 0.7
f = 1, π = 0.6
f = 1, π = 0.3
f =1, π = 0.8
S
5
,2
Oracle:
DBSRV2
*
S
3
,1
Inside
Switch
S
5
,1
Oracle:
DBSRV2
f = 1, π = 0.7
f = 1, π = 0.7
cfg
3
cfg
1
Figure 5: The updated network with PCP added.
most probable path, as we consider a set of most prob-
able paths. Our intuition in doing so stems from the
fact that alternate paths can contribute significantly to
the probability specified by Pen
t
IPN
((s,ℓ),(s
,ℓ
)).
Computing orc. Next, we introduce a simple
sampling-based method that randomly generates se-
quences from (s, ) to (s
,
) of the required length.
These sequences are generated with a probability
proportional to their sequence probability, hence the
computation of orc based on these samples is biased
toward the set of most-probable sequences from (s, )
to (s
,
), which we believe will provide a close
approximation to orc. Our pre-processing method,
ORC SAM below generates a set of sequences that
the attacker could potentially take. Using the se-
quences from SEQ, we can then calculate the penetra-
tion probability (Pen
t
IPN
((s,ℓ),(s
,ℓ
))) by summing over
the probabilities over the sequences in SEQ as op-
posed to summing over all sequences (a potentially
exponential number).
Extensions. Our approach is easily generalizable for
studying other problems related to Cluster-Addition.
For instance, suppose an intruder initiates his infiltra-
tion from one of a set of systems chosen based on a
Algorithm 2: ORC SAM.
Require: Systems s, s
, access levels
s
,
s
, natural num-
bers t, maxIters
Ensure: Set of sequences SEQ
1: curIters = 0, SEQ =
/
0
2: while curIters < maxIters do
3: (s
i
,
i
) = (s,
s
), curSeq = (s,
s
), curLth = 0
4: while curLth < t and (s
i
,
i
) ̸= (s
,
s
) do
5: Select (s
,
) from the set (η
i
×L ) curSeq with
a probability of
f((s
i
,ℓ
i
),(s
,ℓ
))π((s
i
,ℓ
i
),(s
,ℓ
))
(s
j
,ℓ
j
)(η
i
×L )curSeq
f((s
i
,ℓ
i
),(s
j
,ℓ
j
))
6: curSeq = curSeq {(s
,
)}
7: end while
8: If (s
,
) = (s
,
s
) then SEQ = SEQ {curSeq}
9: curIters+ = 1
10: end while
11: Return SEQ
probability distribution. We can encode this problem
in Chain-Addition by adding a dummy system to the
penetration network and establishing relationships to
each of the potential initial systems in the set. The
fitness and conditional probability functions can then
be set up in a manner to reflect the probability distri-
bution over the set of potential initial systems. Note
KeepingIntrudersatLarge-AGraph-theoreticApproachtoReducingtheProbabilityofSuccessfulNetworkIntrusions
25
that this problem would still maintain the same math-
ematical properties and guarantees of our already de-
scribed method to solve the Cluster-Addition prob-
lem. Another related problem that can be solved us-
ing our methods (again with only minor modification)
is to minimize the expected number of compromised
systems in a set of potential targets. As the expected
number of systems would simply be the sum of the
penetration probability for each of those machines, a
simple modification to Line 5 of GREEDY CLUSTER
would allow us to represent this problem, in this case
instead of examining orc we would examine the sum
of the value for orc returned for paths going to each of
the potential targets. Note that by the fact that posi-
tive linear combinations of submodular functions are
also submodular, we are able to retain our theoretical
guarantees here as well.
5 EXPERIMENTAL RESULTS
We conducted several experiments in order to test the
effectiveness of our algorithm under several circum-
stances. As explained in the preceding running ex-
ample we tested our algorithm on a small network
derived from a real network. As shown, by adding
two distraction clusters to that network we were able
to reduce the probability of a successful attack from
8.5% to 2.2%, an almost 75% reduction. While this
result shows relevance to a real-world network and
displays the algorithm on an easily understood scale,
we also wanted to test the scalability of our algorithm
by experimenting on larger networks. Due to the dif-
ficulty of obtaining large datasets for conducting re-
search in cyber security we generated random graphs
that resembled the network topology of the types of
penetration networks we wish to examine, allowing
us to test the scalability of our algorithm. To do this
we generated networks which were divided into lay-
ers, meaning that each system is only connected to
other systems in its own layer or to systems in the
next layer. This is important because it means that in
a network with
n
layers, the shortest path between the
source of the attack and the target is n 1. We did
this because it mimics the topology of the real-world
networks we wished to replicate in our experiments.
It provides structure to our networks so that the sim-
ulated attackers will have to gain access to systems
across a series of layers before gaining access to a
system from which they can access their target. For
all of our experiments we assumed that π = 1. We
did this because it makes it easier to see and under-
stand the relationship between the lower and upper
bounds of the penetration probability. This does not
compromise the value of our results because adding
distraction clusters to a network only effects the rela-
tive fitness of an attack, so the probability of success
for any one attack does not influence our results.
The first test we ran sought to measure the effec-
tiveness of the ORC SAM sampling algorithm by test-
ing against the value of t—the maximum length of
a penetration sequence—and the number of iterations
conducted in the sampling algorithm (ORC SAM). We
found—as would be expected—that the number of it-
erations required for the most accurate possible result
increases as t increases and as the number of system-
level pairs (nodes) in the network increases. We ran
our tests on randomly generated networks with 50 and
100 systems each with 2 levels per system. Figs. 6, 7,
and 8 display the results of this test. Fig. 6 shows that,
as t increases, the number of iterations of ORC SAM
required to get an accurate prediction of the penetra-
tion probability increases. For high values of t, with
a relatively small number of iterations, the difference
between the upper and lower bounds of the test is very
large. This effect is displayed more clearly in Fig. 7
in which this difference is graphed for each value of
t against the number of iterations. Finally, Fig. 8 dis-
plays the time in seconds that each individual test took
to run.
One important factor when implementing distrac-
tion clusters is determining the location in the net-
work in which to place them. The initial instinct may
be that the goal should be to distract the attacker early
in his pursuit and thus the clusters should be placed
close to the source of the attack; however, this has
little effect on the resulting decrease in penetration
probability. We conducted experiments on the same
two networks from the previous test in which we ran-
domly generated a cluster and connected it to 10 sys-
tems in a given layer for each layer in the system
(other than the first and last layers which only have
one system each—the source and the target, respec-
tively). We then measured the changes in penetration
probability that occurred as a result. For each network
we conducted 10 trials. Our results showed little ev-
idence that the proximity to the source of the attack
matters and instead suggest that the size of the layer
is the more influential variable. Clusters with con-
figurations that connected them to systems in a layer
with a relatively small number of systems showed a
larger decrease in the penetration probability. In ad-
dition to being suggested by our evidence, this also
makes intuitive sense because providing the attacker
a distraction in a layer with fewer systems means that
more of his options will lead into the distraction clus-
ter rather than allowing him to progress forward with
his attack.
SECRYPT2014-InternationalConferenceonSecurityandCryptography
26
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
5 15 25 35 45
Penetration Probability Bounds
Iterations (thousands)
t = 6 t = 7 t = 8 t = 9 t = 10 t = 11 t = 12
(a) Network with 50 systems.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
10 30 50 70 90
Penetration Probability Bounds
Iterations (thousands)
t = 7 t = 8 t = 9 t = 10 t = 11
(b) Network with 100 systems.
Figure 6: Upper and lower bounds of the predicted penetration probability vs. number of iterations of ORC SAM, for different
values of t.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
5 15 25 35 45
Penetration Probability Range
Iterations (thousands)
t = 6 t = 7 t = 8 t = 9 t = 10 t = 11 t = 12
(a) Network with 50 systems.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
10 30 50 70 90
Penetration Probability Range
Iterations (thousands)
t = 7 t = 8 t = 9 t = 10 t = 11
(b) Network with 100 systems.
Figure 7: Difference between upper and lower bounds of the predicted penetration probability vs. number of iterations of
ORC SAM, for different values of t.
0
5
10
15
20
25
30
35
40
0 10 20 30 40 50
Execution Time (s)
Iterations (thousands)
t = 6 t = 7 t = 8 t = 9 t = 10 t = 11 t = 12
(a) Network with 50 systems.
0
50
100
150
200
250
300
350
0 20 40 60 80 100
Execution Time (s)
Iterations
t = 7 t = 8 t = 9 t = 10 t = 11
(b) Network with 100 systems.
Figure 8: Execution time vs. number of iterations of ORC SAM, for different values of t.
R² = 0.9437
0.1
0.12
0.14
0.16
0.18
0.2
0.22
0.24
5 7 9 11 13
Penetration Probability
Number of Systems in Layer
Lower Bound Linear (Lower Bound)
(a) Network with 50 systems.
R² = 0.9206
R² = 0.8286
0.015
0.018
0.021
0.024
0.027
0.030
0 5 10 15 20
Penetration Probability
Number of Systems in Layer
Lower Bound Upper Bound Linear (Lower Bound) Linear (Upper Bound)
(b) Network with 100 systems.
Figure 9: Correlation between the number of systems in a layer and the penetration probability when the distraction cluster
was configured to systems in that layers (note that for the network with 50 systems, the bounds nearly matched and only the
lower is displayed).
KeepingIntrudersatLarge-AGraph-theoreticApproachtoReducingtheProbabilityofSuccessfulNetworkIntrusions
27
0
0.05
0.1
0.15
0.2
0.25
0.3
5 15 25 35 45
Penetration Probability
Iterations (thousands)
Without DC With DC
(a) Network with 50 systems, t = 7.
0
0.1
0.2
0.3
0.4
0.5
0.6
5 15 25 35 45
Penetration Probability
Iterations (thousands)
Without DC With DC
(b) Network with 50 systems, t = 8.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
5 15 25 35 45
Penetration Probability
Iterations (thousands)
Without DC With DC
(c) Network with 50 systems, t = 9.
0
0.01
0.02
0.03
0.04
0.05
0.06
10 30 50 70 90
Penetration Probability
Iterations (thousands)
Without DC With DC
(d) Network with 100 systems, t = 7.
0
0.1
0.2
0.3
0.4
0.5
0.6
10 30 50 70 90
Penetration Probability
Iterations (thousands)
Without DC With DC
(e) Network with 100 systems, t = 8.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
10 30 50 70 90
Penetration Probability
Iterations (thousands)
Without DC With DC
(f) Network with 100 systems, t = 9.
Figure 10: Penetration probabilities before and after adding the cluster chosen by GREEDY CLUSTER.
Fig. 9 shows the results from the tests. The y-axis
represents the average penetration probability with
the configuration added to the layer while the x-axis
represents the number of systems in the layer. The
dotted line shows the result of the linear regression
which shows a linear relationship between the two
variables. While we show a linear regression the rela-
tionship is not exactly linear as the effect is lessened
as the layers become larger. Note that the average
percent decrease in penetration probability when clus-
ters were added to the network with 50 systems was
20.7% while the average when they were added to the
larger network was only 6.4%. We hope to do further
testing in the future to identify other variables that af-
fect the outcome. In addition to investigating the ideal
location for placing a cluster we also examined the
effects of the number of iterations of ORC SAM that
are run on the results of the test. These results are
shown in Fig. 10. As explained earlier, ORC SAM is
a heuristic that generates a random sample of possi-
ble paths between the source and target. To test this
we ran GREEDY CLUSTER with varying numbers of
iterations run on ORC SAM. We generated ten possi-
ble configurations for GREEDY CLUSTER to choose
from and assigned it to choose the optimal configu-
ration. We ran the test on both networks with itera-
tions of ORC SAM ranging from 5,000 to 50,000 in
increments of 5,000 on the smaller network and from
10,000 to 100,000 in increments of 10,000 on the
larger one. We ran this test with the values of t—the
maximum length of a penetration sequence—set to 7,
8, and 9 for three different tests on both networks.
We found that regardless of the number of iterations
run or the value of t, GREEDY CLUSTER selected the
same configuration and there was a small difference
between the percent decrease in penetration probabil-
ities after the clusters were added to the network. This
suggests that accurate predictions about the effects of
a cluster on a network can be made without sampling
all possible paths from the source to the target. Ad-
ditionally, as stated earlier, it suggests that seeking to
reduce the probability of the attack being reached in
the shortest number of steps will also help reduce the
probability of paths of greater lengths.
6 CONCLUSIONS
Despite significant progress in the area of intrusion
prevention, it is well known that not all intrusions
can be prevented, and additional lines of defense are
needed in order to cope with attackers capable of cir-
cumventing existing intrusion prevention systems.
In this paper, we proposed a novel approach to in-
trusion prevention that aims at delaying an intrusion,
rather than trying to stop it, so as to control the proba-
bility that an intruder will reach a certain goal within
a specified amount of time. In the future, we plan to
do more experiments to better understand the factors
that influence the ideal location of distraction clus-
SECRYPT2014-InternationalConferenceonSecurityandCryptography
28
ters and to determine the relationship between pen-
etration probability, iterations of ORC SAM, and the
selection of configurations for adding distraction clus-
ters as well as investigate the impact of the length of
distraction clusters on the network and the location of
the point at which the distraction clusters reconnect to
the network.
REFERENCES
Abbasi, F., Harris, R., Moretti, G., Haider, A., and An-
war, N. (2012). Classification of malicious network
streams using honeynets. In Global Communications
Conference (GLOBECOM), pages 891–897.
Alpcan, T. and Baar, T. (2010). Network Security: A Deci-
sion and Game-Theoretic Approach. Cambridge Uni-
versity Press, New York, NY, USA, 1st edition.
Chen, C.-M., Cheng, S.-T., and Zeng, R.-Y. (2013). A
proactive approach to intrusion detection and malware
collection. Security and Communication Networks,
6(7):844–853.
Chen, W., Wang, C., and Wang, Y. (2010). Scalable in-
fluence maximization for prevalent viral marketing in
large-scale social networks. In Proceedings of the 16th
ACM SIGKDD international conference on Knowl-
edge discovery and data mining, pages 1029–1038.
Evans, D., Nguyen-Tuong, A., and Knight, J. C. (2011).
Moving Target Defense: Creating Asymmetric Un-
certainty for Cyber Threats, chapter Effectiveness of
Moving Target Defenses, pages 29–48. Springer.
Feige, U. (1998). A threshold of ln n for approximating set
cover. J. ACM, 45(4):634–652.
Jajodia, S., Ghosh, A. K., Subrahmanian, V. S., Swarup, V.,
Wang, C., and Wang, X. S., editors (2013). Moving
Target Defense II: Application of Game Theory and
Adversarial Modeling, volume 100 of Advances in In-
formation Security. Springer, 1st edition.
Jajodia, S., Ghosh, A. K., Swarup, V., Wang, C., and Wang,
X. S., editors (2011). Moving Target Defense: Cre-
ating Asymmetric Uncertainty for Cyber Threats, vol-
ume 54 of Advances in Information Security. Springer.
Manadhata, P. K. and Wing, J. M. (2011). An attack surface
metric. IEEE Transactions on Software Engineering,
37(3):371–386.
Nemhauser, G. L., Wolsey, L. A., and Fisher, M. (1978).
An analysis of approximations for maximizing sub-
modular set functionsi. Mathematical Programming,
14(1):265–294.
P
´
ıbil, R., Lis
´
y, V., Kiekintveld, C., Bosansk
´
y, B., and Pe-
choucek, M. (2012). Game theoretic model of strate-
gic honeypot selection in computer networks. In
GameSec, pages 201–220.
Shakarian, P., Shakarian, J., and Ruef, A. (2013). Introduc-
tion to Cyber-Warfare: A Multidisciplinary Approach.
Syngress.
Sweeney, P. and Cybenko, G. (2012). An analytic approach
to cyber adversarial dynamics. In SPIE Defense, Se-
curity, and Sensing, pages 835906–835906. Interna-
tional Society for Optics and Photonics.
Williamson, S. A., Varakantham, P., Hui, O. C., and Gao,
D. (2012). Active malware analysis using stochastic
games. In Proceedings of the 11th International Con-
ference on Autonomous Agents and Multiagent Sys-
tems - Volume 1, AAMAS ’12, pages 29–36, Rich-
land, SC. International Foundation for Autonomous
Agents and Multiagent Systems.
APPENDIX
Proof of Theorem 1. The Cluster Addition Problem is NP-
hard and the associated decision problem is NP-Complete
when the number of sequences from (s,) to (s
,
) is a poly-
nomial in the number of nodes in the intruder penetration
network.
Proof. Membership in NP (if the number of sequences from
(s,) to (s
,
) is polynomial can be shown when the certifi-
cate is the set of configuration-cluster pairs.
For NP-hardness, consider the set cover problem (Feige,
1998) where the input consists of as set of elements S,
a family of subsets of S denoted as H, and natural num-
bers K, X. The output of this problem is a subset of H
of size K or less such that their union covers X or more
elements of S. This problem is NP-hard and can be em-
bedded into an instance of the cluster addition problem
as follows: L = {0, 1}, S = {s,t} {v
w
|w S}, R =
{(s,v
w
),(v
w
,t)|w S}, w S, set π((s, 1),(v
w
,1)) = 1 and
π((v
w
,1), (t,1)) = 1 (otherwise set π by definition), w S,
set f((s,1),(v
w
,1)) = 1 and f((v
w
,1), (t, 1)) = 1 (otherwise
set f by definition),
s
,
s
= 1, CL = {cl}, CFG = {cfg
h
|h
H}, for each cfg
h
CFG and cl set π
cfg
h
,cl
(v
w
) = 1 if s h
and 0 otherwise, for each cfg
h
CFG and cl set f
cfg
h
,cl
(v
w
) =
|S| if s h and 0 otherwise, x = 1
|S|−X
|S|
X
|S|(|S|+1)
, k = K,
and t = 2. Clearly this construction can be completed in
polynomial time.
Next, we show that a solution to set cover will provide
a solution to the constructed cluster-addition problem. If
H
is a solution to set cover, select the set {cfg
h
|h H
}.
Clearly this meets the cardinality constraint. Note that in
the construction, all sequences from (s,1) to (t, 1) are of the
form (s, 1),(v
w
,1), (t,1) where w S. Note that as every
system v
w
S { s,t} is now connected to a cluster. Hence,
each cluster now has had its probability reduced from 1/|S|
to at most 1/(|S |(|S| + 1). Hence, Pen
t
IPNPCP
((s,ℓ),(s
,ℓ
)) <
|S|−X
|S|
+
X
|S|(|S|+1)
which completes this claim of the proof.
Going the other way, we show that a solution to the
constructed cluster-addition problem will provide a solu-
tion to set cover. Given cluster-addition solution PCP, con-
sider H
= {h|(cfg
h
,cl) PCP}. Note that, by the construc-
tion, all elements of PCP are of the form (cfg
h
,cl) where
h H
. Clearly, the cardinality constraint is met by the
construction. Suppose, BWOC, H
is not a valid solution
to set cover. We note that this must imply that there are
some v
s
that are not attached to a distraction cluster. Let
us assume there are δ number of these systems. Hence,
Pen
t
IPNPCP
((s,ℓ),(s
,ℓ
)) >
|S|−X+δ
|S|
+
Xδ
|S|(k|S|+1)
. Let us now
assume, by way of contradiction, that this quantity is less
than or equal to
|S|−X
|S|
+
X
|S|(|S|+1)
, which is the upper bound
on Pen
t
IPNPCP
((s,ℓ),(s
,ℓ
)) is at least X of the v
s
systems have
KeepingIntrudersatLarge-AGraph-theoreticApproachtoReducingtheProbabilityofSuccessfulNetworkIntrusions
29
a distraction cluster:
|S| X + δ
|S|
+
X δ
|S|(k|S| + 1)
|S| X
|S|
+
X
|S|(|S| + 1)
(8)
δ(k|S| + 1) + X δ
|S|(k|S| + 1)
X
|S|(|S| + 1)
(9)
(δk|S| + X)(|S| + 1) X(k |S| + 1) (10)
However, as δk|S|+X > k|S|+ 1 and as |S|+1 > X , this
give us a contradiction, completing the proof.
Proof of Theorem 2. With a cardinality constraint, find-
ing set PCP s.t. orc(PCP) cannot be approximated in PTIME
within a ratio of
e1
e
+ε for some ε > 0 (where e is the base
of the natural log) unless P=NP.
Proof. Follows directly from the construction of Theorem
1 and Theorem 5.3 of (Feige, 1998). Slightly longer
version below: Note that Theorem 1 shows that we can
find an exact solution to set-cover in polynomial time using
an instance of the Cluster Addition problem. A optimiza-
tion variant of this problem, the MAX-K-COVER prob-
lem (Feige, 1998) takes the same input and returns a subset
of H size K whose union covers the maximum number of
elements in S. This variant of the problem cannot be ap-
proximated within a ratio of α =
e1
e
+ ε for some ε > 0 by
Theorem 5.3 of (Feige, 1998).
Proof of Lemma 1. For PCP
PCP CFG ×CL, orc(PCP
)
orc(PCP).
Proof. Given an instance of the cluster addition problem,
consider a single sequence from s to s
. Clearly the prob-
ability associated with that sequence will either remain the
same with the addition of any configuration-cluster pair to
the penetration network as the denominator of the proba-
bility of transitioning between system-level pairs will only
increase (due to the additive nature of fitness in the denom-
inator). Likewise, subsequent configuration-cluster pair ad-
ditions will lead to further decrease. Hence, the penetra-
tion probability monotonically decreases with the addition
of configuration-cluster pairs as it is simply the sum of the
sequence probabilities form s to s
of length t which makes
orc monotonically increasing, as per the statement.
Proof of Lemma 2. For PCP
PCP CFG × CL and pc =
(cfg,cl) / PCP we have:
orc(PCP {pc}) orc(PCP) orc(PCP
{pc}) orc(PCP
)
Proof. It is well known that a positive, linear combination
of submodular functions is also submodular. Hence,
without loss of generality, we can prove the statement by
only considering a single sequence. Let lth be the length
(number of transitions) of the sequence. For the ith (s,) in
the sequence, let F
i+1
=
(s
,ℓ
)η
i
×L
f((s,), (s
,
)),
B
i+1
=
(s
,ℓ
)PCP
×L
f((s,), (s
,
)), D
i+1
=
(s
,ℓ
)PCP×L
f((s,), (s
,
)) B
i+1
, E
i+1
= f((s, ), pc).
Using this notation, we can apply the defini-
tion of orc and some easy algebra, we obtain that
i
1
F
i
+B
i
+D
i
i
1
F
i
+B
i
+D
i
+E
i
is less than or equal to
i
1
F
i
+B
i
i
1
F
i
+B
i
+E
i
. Suppose, BWOC, the statement is
false, this implies that
i
F
i
+B
i
F
i
+B
i
+E
i
(
1
i
F
i
+B
i
+E
i
F
i
+B
i
+D
i
+E
i
)
is
greater than 1
i
F
i
+B
i
F
i
+B
i
+D
i
. Here, the term
i
F
i
+B
i
F
i
+B
i
+E
i
decreases as each of the E
i
s increase and that this value is
no more than 1. Hence, the following must be true (under
the assumption that the original statement is false).
i
F
i
+ B
i
+ E
i
F
i
+ B
i
+ D
i
+ E
i
<
i
F
i
+ B
i
F
i
+ B
i
+ D
i
For the above statement to be true, there must exist
at least one i such that (F
i
+ B
i
+ E
i
)(F
i
+ B
i
+ D
i
) <
(F
i
+ B
i
)(F
i
+ B
i
+ D
i
+ E
i
) which implies D
i
,E
i
0, thus
leading us to a contradiction and completing the proof.
Proof of Theorem 3. GREEDY CLUSTER provides the
best PTIME approximation of orc unless P=NP if orc can be
solved in PTIME.
Proof. If orc can be solved in PTIME, it is easy to then
show that GREEDY CLUSTER runs in PTIME. The results
of (Nemhauser et al., 1978) show a greedy algorithm pro-
vides a
e
e1
approximation for the maximization of a non-
decreasing submodular function that returns zero on the
empty set. Clearly, orc(
/
0) = 0 and we showed that it is non-
decreasing and submodular in Lemmas 1 and 2 respectively
- which means that GREEDY CLUSTER provides a
e
e1
ap-
proximation to the maximization of orc. This matches the
theoretical bound proved in Theorem 2 which holds unless
P=NP.
SECRYPT2014-InternationalConferenceonSecurityandCryptography
30