Professional Documents
Culture Documents
Defeating Terrorist
Networks with
Game Theory
Tomasz P. Michalak, University of Oxford
Talal Rahwan, Masdar Institute of Science and Technology
Oskar Skibski, Kyushu University
Michael Wooldridge, University of Oxford
D
espite enormous efforts to tackle terrorism will depend on the application domain, but in the
in the aftermath of 9/11, many terrorist net- analysis of terrorist networks, nodes are interpreted
as members of the network, edges indicate links be-
works are growing in size and in threat. Faced with
tween them, and weights indicate the strength of the
this challenge, security agencies urgently require link (for example, the number of meetings between
robust and efficient techniques to analyze and the two individuals). Centrality analysis aims to cre-
understand the structure and operation of terror- ate a principled ranking of the importance of nodes
ist organizations. One key problem is that of iden- within such a network. Since “importance” depends
tifying the key members of the organization using on the problem at hand, researchers have consid-
information about the terrorist network’s topol- ered many different centrality measures. Among
ogy: this capability would enable security agencies them, degree centrality, closeness centrality, and be-
to focus severely limited resources on just those key tweenness centrality are the most well-known and
members. widely applied. According to weighted degree cen-
To this end, many standard measures of centrality trality, a node’s importance is equal to the weight
from the field of social network analysis can be of that node’s adjacent edges (that is, the weighted
used. Centrality measures aim to give a numerical degree of the node). For instance, in the network in
characterization of a node’s significance in a Figure 1, nodes v1 and v2 are the most important ac-
network. For example, IBM’s Analyst Notebook— cording to weighted degree centrality, because they
a software package used worldwide by law have a weighted degree of 6 each, which is greater
enforcement and intelligence agencies—supports than that of any other node in the network. In con-
the degree, closeness, and betweenness centralities, trast, according to closeness centrality, a node’s im-
which are probably the most widely used centrality portance is based on the average distance between
measures. However, these measures have well- that node and other nodes in the network. In Fig-
known limitations, and the development of more ure 1, for example, this measure considers node v8
sophisticated measures is currently a key research to be the most important in the network. Finally,
area. One important thread of current research is to betweenness centrality focuses on shortest paths
apply techniques from the field of cooperative game between any two nodes in the network; the more
theory to this problem. In this article, we present a shortest paths a given node belongs to, the higher
survey of this work. its ranking. In Figure 1, node v11 is the most impor-
tant from the perspective of betweenness centrality.
Classic Measures of Centrality All of these centralities can be defined on weighted,
The networks we are interested in are undirected unweighted, directed, and undirected networks. For
weighted graphs: they consist of a set of nodes instance, the unweighted in-degree centrality of a
(vertices) and undirected edges that connect some node counts the total number of its incoming edges.
nodes; each edge is associated with a numerical Clearly, all three measures expose a node’s
weight. The specific interpretation of the network different characteristics. For example, assume that
A
cooperative game models scenarios where individuals, The Shapley value of i is a weighted average of player i’s
or players, benefit from cooperation, and where bind- marginal contributions, where the marginal contribution
ing agreements are possible. This means it is possible of i to a coalition C is v (C ∪ {i }) − v(C). More precisely, for
for players to agree on which groups, or coalitions, to form, any permutation of players, let us consider the marginal
and on how the payoff, or value, of the formed coalitions contribution of vi to the coalition consisting of all players
shall be divided among the members. A coalition’s value is that appear before it in the permutation. Then, the Shapley
often assumed to depend entirely on the identities of the co- value is simply the average such marginal contribution taken
alition members. In such games, the function that assigns a over all permutations.
value to every coalition is called the characteristic function, Myerson followed a similar approach, but for games
and is denoted by v. defined on a graph G in which nodes represent players and
The vector that specifies each player’s share is called the edges represent communication channels.1 Thus, every
payoff vector. The main question here is which payoff vector coalition can be thought of as a subgraph of G. Myerson
should we adopt? One common way of addressing this assumed that if a coalition, or a subgraph, C was not
question is to carefully select a number of desirable properties, connected in G, then the value of C is simply the sum of the
and then prove that there exists only one possible payoff vector values of the connected components in C. The intuition is that
that satisfies all of those properties. The selected properties are disconnected components cannot communicate with one
then called the axioms of this unique vector. In this context, another, and so there is no value added when considering
Shapley focused on scenarios where the players form the grand their union. Formally, the value of C is
coalition—that is, one big coalition containing all players in the
game.1 He proposed the following axioms:
⎧ v (C )
⎪ if C is connected in G
t Efficiency. The grand coalition value is divided entirely w (C ) = ⎨ ∑ v (S )
otherwise.
among the players (that is, there are no leftovers). ⎩⎪connected component S in C
t Null player. The share is zero for every “null player”—that
is, a player whose membership in any coalition does not
affect the value of that coalition. For such scenarios, Myerson proposed the following
t Symmetry. All “symmetric” players receive equal shares, axioms:
where any two players are symmetric if replacing one with t Component efficiency. The payoff of every connected
the other in any coalition makes no impact on that coali- component is divided entirely among its members (that is,
tion’s value. there are no leftovers).
t Additivity. Given two games, G1 and G2, that involve the t Fairness. Any two connected players benefit equally from
same set of players, and given a third game, G3, in which the bilateral connection between them.
a coalition’s value is the sum of its value in G1 and in G2,
the share of a player in G3 should be the sum of its share Myerson proved that there is a unique payoff vector
in G1 and in G2. satisfying the above axioms, whereby the share of a player i,
called the Myerson value of i, is
Shapley proved that there is a unique payoff vector satis-
fying all of the above, whereby the share of a player i, called
v w
“the Shapley value of i,” is Myerson _ valuei = Shapley _ valuei .
C !( n − C − 1) ! Reference
v
Shapley _ valuei = ∑ n!
(v (C ∪ {i }) − v (C )) . 1. M. Maschler, E. Solan, and S. Zamir, Game Theory, Cambridge
C ⊆ N \ {i } Univ. Press, 2013.
For instance, the value of the The fraction in the above character- those authors focused on the Shapley
connected coalition {v1, v5} in Figure istic function was proposed by Linde- value, Skibski and colleagues showed
1 is 2, and the value of the connected lauf and colleagues, who provided the that the Myerson value has more in-
coalition {v2, v3, v9, v10} is 2/3. In following rationale: “A terrorist orga- teresting properties as a centrality
contrast, the coalition {v1, v2, v3, v5, nization will try to shield its impor- measure.2
v9, v10} is disconnected, meaning tant players by keeping the frequency
that the terrorists involved cannot and duration of their interaction with Example 3: Centrality with Skills
communicate, and so they will proceed others to a minimum. However, to be Let us again consider the situation
with their activities as two independent able to coordinate and control the at- when the critical skills of nodes v1,
groups. Consequently, our graph- tack, an important player needs to v2, v6, and v11 are known to the in-
restricted game assigns the value of 2 maintain relationships with other in- vestigators. Michalak and colleagues
+ 2/3 to {v1, v2, v3, v5, v9, v10}. dividuals in the network.”3 Although proposed the Myerson value-based
A
n interesting property of game-theoretic centrality is t The solution concept. This concept specifies the criteria the
its flexibility. Specifically, it can be adapted to a partic- players use to decide which coalitions to form and which
ular application, along three general dimensions. payoff vector to adopt. Different solution concepts are based
on different prescriptive and normative considerations. Ex-
t The form of the game. The game is usually assumed to be in
amples include the Shapley value and Myerson value (see the
characteristic function form—that is, a coalition’s value de-
“Coalitional Games” sidebar), which focus on fairness, and
pends solely on the identities of its members. However, other
the core and the nucleolus, which focus on stability.2
forms may better model the application at hand. For in-
stance, partition function games allow for a coalition’s value A recent overview on the literature on game-theoretic
to be influenced by the way nonmembers are partitioned. Al- centrality can be found in Tarkowski and colleagues.3
ternatively, generalized characteristic function games assume
that the members of a coalition are ordered, and allow for a
coalition’s value to be influenced by the order of its members. References
1. P.T. Michalak et al., “Efficient Computation of the Shapley Value
t The coalition-value function. Usually, an arbitrary coali-
for Game-Theoretic Network Centrality,” J. Artificial Intelligence
tion-value function is assumed. Thus, one can pick the Research, vol. 46, 2013, pp. 607–650.
function that best represents the worth of a coalition 2. M. Maschler, E. Solan, and S. Zamir, Game Theory, Cambridge
given the application at hand. For instance, Michalak and Univ. Press, 2013.
colleagues consider alternative such functions when study- 3. M. Tarkowski et al., “Game-Theoretic Network Centrality
ing the problem of information diffusion in networks.1 Measures: A Critical Survey,” mimeo, Univ. of Oxford, 2014.
centrality measure that accounts for These examples demonstrate that The Computational Challenge
such qualitative data.4 Specifically, game-theoretic centrality measures Unfortunately, due to their inher-
they proposed the following charac- account for individual properties of ent combinatorial nature, comput-
teristic function, where F denotes a nodes and edges, expressed both quan- ing solution concepts for coalitional
terrorist capable of funding multiple titatively and qualitatively (property games is often computationally chal-
attacks, B denotes a terrorist capable 1). They also analyze all groups (that lenging. In fact, already the size of the
of handling bombs, and M denotes a is, subsets) of nodes (property 2). input (that is, the number of elements
terrorist skilled in martial arts: Finally, they use all those pieces of of the characteristic function) grows
information to construct a coherent exponentially with the number of
value(C)
ranking of individual nodes (prop- players. Luckily, in many cases the
⎧ ⎧ M ⎫ erty 3). On top of this, by drawing coalition values in the coalitional
⎪min ⎨the number of B, 2 in C ⎬
⎪⎪ ⎩ ⎭ parallels between networks and co- game have some underlying structure,
= ⎨ × ( the number of F in C ) if C is connected, alitional games, we can build on a which makes it possible to represent
⎪ the sum of values of the large body of literature on game-the- the characteristic function compactly.
⎪
⎪⎩ connected componentss of C otherwise. oretic solution concepts. Importantly, This has led to an important research
the game-theoretic approach to cen- direction in algorithmic game theory,
Intuitively, this qualitative charac- trality places no assumptions on the aimed at identifying classes of games
teristic function represents scenarios in coalition-evaluation function, nor the that admit such a compact represen-
which every attack requires a terrorist adapted solution concept, nor even tation, and determining whether and
with the necessary funding, in addition the form of the game (see the “Flex- how certain solution concepts are
to one bomb expert and two terrorists ibility of Game-Theoretic Central- computable in polynomial time.
skilled in martial arts. This can be gener- ity” sidebar); these can be adjusted as In the context of game-theoretic cen-
alized to more than just a single attack— necessary, thus providing a desirable trality, the underlying coalitional game is
for example, the terrorist providing the degree of flexibility to capture a wide defined over the network, and is some-
funds, combined with two bomb experts range of scenarios. how influenced by the network topol-
and four martial artists, can carry out But alas, every rose has its thorns. ogy. This suggests that, in some cases,
two attacks. Naturally, the numbers and All the advantages of game-theoretic such coalitional games can be (but do
skills specified in the above function can centrality measures might never ma- not have to be) represented compactly.
be adjusted to better reflect the scenario terialize if we cannot compute those For instance, this is the case for the
under investigation. Note that, for each measures in the first place. Next, we games in Examples 1, 2, and 3 given ear-
attack, the group must be connected to discuss recent works that address lier. However, having a compact repre-
be able to coordinate its activities. this issue. sentation does not necessarily imply that
υ19 υ2 - S. Suqami
M υ3 - Waleed Alshehri
way it traversed the graph; enumerat- υ18
υ4 - A. A. AI-Omari
ing all induced connected subgraphs
υ17 υ5 - M. Atta
required breadth-first search (BFS), M
whereas identifying cut vertices re- υ6 - M. Moqed
υ14
quired depth-first search (DFS). υ7 - M. AI-Shehhi
To address this issue, Skibski and υ15 υ8 - F. Ahmed
υ16 M
colleagues proposed the first algo- υ9 - H. Hanjour
M υ13
rithm that performs both opera- υ10 υ10 - K. AI-Midhdar
tions simultaneously. 2 The main M υ11 - S. Alhazmi
υ11
trick is to perform the first operation υ12 - Z. Jarrah
P
(enumerating all induced connected υ13 - A. AI-Haznawi
subgraphs) using DFS rather than BFS, υ9 υ12 υ14 - S. Alghamdi
thus making it compatible with the M υ8
υ15 - M. Alshehri
other operation’s performance (iden- P
υ16 - N. Alhazmi
tifying cut vertices). Broadly speak- υ7 υ17 - H. Alghamdi
ing, the DFS enumeration algorithm
υ6 υ18 - A. Alnaml
starts with a subgraph consisting of P
υ19 - A. Alghamdi
a single node, and gradually expands υ5
M
the subgraph to (some of) its neigh-
P
boring nodes in a DFS manner. This is
done using a divide-and-conquer tech-
nique, which works as follows: when- υ4
ever a neighbor is found that has not
yet been visited by the algorithm, the M
calculations are split into two parts,
the first with the neighbor added to
υ3
the subgraph, and the second with the
neighbor marked as forbidden. This
υ1
entire process is repeated recursively
such that eventually all induced sub- υ2
graphs are enumerated without any
M
redundancies (that is, without per-
forming an operation twice). For a
Figure 2. World Trade Center 9/11 terrorist network. 3,6 The letter P stands for a
detailed description of the algorithm,
pilot, and the letter M indicates a terrorist skilled in martial arts.
as well as the pseudocode, see Oskar
Skibski and colleagues2
Case Study: World Trade to split into four groups. Each group assumed that two terrorists with such
Center, 9/11 had to contain at least one terrorist skills were needed per flight.
In the already classic work, Valdis E. with pilot training (P) and some ter- Let us now consider a qualita-
Krebs constructed the World Trade rorists skilled in martial arts (M) in tive characteristic function similar to
Center 9/11 terrorist network from order to take control over the crew that of Example 3, but now modified
publicly available sources and used and the cabin. To fulfill these require- such that a P-terrorist (pilot) and two
standard centrality measures to de- ments, terrorists S. Suqami and A.A. M-terrorists (martial arts) are needed
termine the key terrorists.6 Let us fo- Al-Omari took part in martial arts to carry out a single attack. Table 1
cus on the core of the 9/11 network training in the weeks before the attack; compares the (qualitative) Myerson-
composed of 19 hijackers with 32 it is unknown who else was skilled in value-based centrality with the other
relationships, as illustrated in Fig- martial arts. Therefore, we assigned standard centralities for Figure 2. We
ure 2. To hijack the four planes on 11 this skill to seven other terrorists observe that, as expected, the new cen-
September 2001, the terrorists needed chosen uniformly at random, and trality measure promotes the terrorist
with the necessary skills. Interest- account for the properties of each ter- Acknowledgments
ingly, the M-terrorist A.A. Al-Omari is rorist, compute the synergies in vari- Tomasz Michalak and Michael Wooldridge
ranked higher than the P-terrorist M. ous groups (that is, subsets) of those were supported by the European Research
Atta. In other words, when consider- terrorists, and aggregate all this in- Council under Advanced Grant 291528
(“RACE”). Oskar Skibski and Tomasz
ing the positions (in the network) and formation into a coherent ranking Michalak were supported by the Polish
the skills of different terrorists, it turns of individual terrorists. This was not National Science grant DEC-2013/09/D/
out that the martial arts skill of A.A. possible until recently, when game- ST6/03920.
Al-Omari is more often indispensable theoretic centrality measures were
to increasing the attack capabilities proposed. We showed how these new References
of connected coalitions of terrorists measures capitalize on decades of re- 1. M.G. Everett and S.P. Borgatti, “The
than the pilot skills of M. Atta. Con- search in cooperative game theory, Centrality of Groups and Classes,”
versely, N. Alhazmi—the terrorist with thus providing a rich pool of solu- J. Mathematical Sociology, vol. 23, no. 3,
no skills, who is ranked at the top by tion concepts to evaluate individuals 1999, pp. 181–201.
all three standard centrality measures on the basis of their performance in 2. O. Skibski et al., “Algorithms for
(unable to recognize skills)—is now different groups. Unfortunately, these the Shapley and Myerson Values in
ranked ninth. new measures are inherently com- Graph-Restricted Games,” Proc. 13th
plex and raise various computational Int’l Conf. Autonomous Agents and
challenges, even for relatively small
T
Multiagent Systems (AAMAS 2014),
he problem of identifying key networks. However, we showed how 2014, pp. 197–204.
members of a terrorist network is to overcome some of these computa- 3. R. Lindelauf, H. Hamers, and B.
an important topic in the literature; tional challenges, following recent re- Husslage, “Cooperative Game Theoretic
it has attracted significant interest sults from algorithmic game theory. Centrality Analysis of Terrorist Networks:
within the social network analysis Hopefully, the further development The Cases of Jemaah Islamiyah and
community and well beyond. In this of these measures will shed more light Al Qaeda,” European J. Operational
article, we argued that this problem on the dark nature of covert terrorist Research, vol. 229, no. 1, 2013,
requires a centrality measure that can organizations. pp. 230–238.
EXECUTIVE STAFF
Executive Director: Angela R. Burgess; Associate Executive Director &
Director, Governance: Anne Marie Kelly; Director, Finance & Accounting:
PURPOSE: The IEEE Computer Society is the world’s largest association John Miller; Director, Information Technology & Services: Ray Kahn;
of computing professionals and is the leading provider of technical Director, Membership Development: Eric Berkowitz; Director, Products &
information in the field. Services: Evan Butterfield; Director, Sales & Marketing: Chris Jensen
MEMBERSHIP: Members receive the monthly magazine Computer,
discounts, and opportunities to serve (all activities are led by volunteer COMPUTER SOCIETY OFFICES
members). Membership is open to all IEEE members, affiliate society Washington, D.C.: 2001 L St., Ste. 700, Washington, D.C. 20036-4928
members, and others interested in the computer field. Phone: GFax: +1 202 728 9614
COMPUTER SOCIETY WEBSITE: www.computer.org Email: hq.ofc@computer.org
Los Alamitos: 10662 Los Vaqueros Circle, Los Alamitos, CA 90720
Next Board Meeting: 26–30 January 2015, Long Beach, CA, USA Phone: GEmail: help@computer.org