European Journal of Operational Research 283 (2020) 797–811

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Invited Review

A survey of network interdiction models and algorithms

J. Cole Smith∗, Yongjia Song
Department of Industrial Engineering, Clemson University, United States

a r t i c l e i n f o a b s t r a c t

Article history: This paper discusses the development of interdiction optimization models and algorithms, with an em-
Received 24 December 2018 phasis on mathematical programming techniques and future research challenges in the field. After pre-
Accepted 13 June 2019
senting basic interdiction concepts and notation, we recount the motivation and models behind founding
Available online 18 June 2019
research in the network interdiction field. Next, we examine some of the most common means of solv-
Keywords: ing interdiction problems, focusing on dualization models and extended formulations solvable by row-
Networks generation techniques. We then examine contemporary interdiction problems involving incomplete in-
Network interdiction formation, information asymmetry, stochasticity, and dynamic play. We conclude by discussing several
Min-max games emerging applications in the field of network interdiction.
Mixed integer programming formulations
Decomposition algorithms © 2019 Elsevier B.V. All rights reserved.

1. Introduction optimization models in Section 1.3; and previewing the remainder

of this paper in Section 1.4.
In the defense context, interdiction refers to actions that serve
to block or otherwise inhibit an adversary’s operations, and of- 1.1. Basic interdiction setup
ten regards attacks against supply chain operations or communi-
cations. The optimization literature regards interdiction problems Many interdiction studies in the literature make a common set
as leader–follower games in which the leader takes interdiction of key assumptions regarding the game being examined. Some of
actions to maximize the minimum objective that a follower can these are enumerated below.
obtain in solving its optimization problem. The leader’s interdic-
tion actions can impact the follower’s objective, feasible region, or A1. All problem data is known to both the leader and the fol-
both. (Interdiction also refers to a leader’s minimization of a fol- lower.
lower’s maximization problem, such as in the interdiction of max- A2. The leader is certain of the effect that its interdiction actions
imum flow problems.) Because of the origins of interdiction study, have on the follower’s problem.
most such problems in the literature focus on network interdiction. A3. The leader and follower play a zero-sum game in the sense
Numerous variations on interdiction problems have appeared that the value of the game is given by the follower’s objec-
in the literature in the past few decades, due both to the rapid tive, with the leader seeking to maximize (or minimize) the
growth in employing these models in practical settings and to the minimum (or maximum) value that the follower can achieve
number of theoretical and algorithmic innovations in solving them. via constrained optimization.
The purpose of this paper is to give an introduction to interdiction A4. In each round of the interdiction game, the leader and fol-
problems along with a brief history of their development, before lower each make one set of decisions, with the leader mak-
exploring the next generation of interdiction problems and their ing all of its decisions before the follower makes its set of
solution approaches. decisions. (This setup is often referred to as a “defender–
We begin this paper by describing fundamental assumptions attacker” or “attacker–defender” game, depending on the
pertaining to interdiction problems in Section 1.1; providing ba- context of the game.)
sic interdiction formulations and terminology for these problems A5. Only one round of the game is played.
in Section 1.2; comparing interdiction models to bilevel and robust
These problems belong to the more general class of Stackelberg
games (von Stackelberg, 1952), which involve a single leader and
one or more followers, not necessarily playing a zero-sum game. As
∗
Corresponding author. in assumption A4, in a Stackelberg game the leader makes its move
E-mail addresses: [email protected] (J.C. Smith), [email protected] first, and then the follower(s) make their decisions in response to
(Y. Song). the leader’s decisions.

https://doi.org/10.1016/j.ejor.2019.06.024
0377-2217/© 2019 Elsevier B.V. All rights reserved.
798 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

4 In addition, as we discuss later in Section 4, there also exist sev-

2 5 eral recent variations of interdiction problems in which assump-
tions A1–A5 do not all hold.
9 8(+2) 10
1.2. Interdiction formulations
1 2 5 6
We adopt the convention that the leader’s variables are denoted
11 by variables x ∈ RnL (where RnL is the set of nL -dimensional real
vectors) and the follower’s variables by y ∈ RnF , where nL and nF
3(+2) 4(+3)
are both positive integers. Letting the function (x) represent the
3 4 value of interdiction decision x, a general form of the interdiction
12 problem is given by:
Fig. 1. An illustration of the shortest path interdiction problem. max (x ) (1a)

s.t. x ∈ X, (1b)
Example 1. For instance, one well-studied class of such problems
is the shortest path interdiction problem (Israeli & Wood, 2002). where X is a non-empty (and usually bounded) set in R nL . The
Here, a leader attacks arcs with the goal of maximizing the short- value of interdiction decision x, (x), is defined as:
est path cost (distance) that the follower can obtain. An arc that (x ) = min f (x, y ) (2a)
is attacked has its cost increased from some nominal value. (The
leader is constrained by the number of arcs it can attack, or else s.t. y ∈ Y ( x ), (2b)
there exists a trivial optimal solution in which all arcs are inter-
where f(x, y) represents the follower’s objective function (affected
dicted.) By assumption A1, the leader knows the follower’s ori-
by the leader’s action x). We define Y (x ) ⊆ RnF as the set of fea-
gin and destination nodes, along with the nominal arc costs and
sible actions for the follower given any leader’s action x ∈ X. Con-
the increased values of those arc costs if attacked. Assumption A2
densed, this model becomes simply
guarantees that the leader’s attacks will lengthen the targeted arcs  
as anticipated. Assumption A3 assures that both players agree on
the objective and that they play a max-min game. Assumption A4 z = max min { f (x, y )} .
∗
(3)
x∈X y∈Y (x )
states that the leader must commit all of its attacks before the fol-
lower chooses a path, and that the follower’s path choice is the To illustrate this general form, consider the shortest path inter-
only decision that the follower makes. Assumption A5 states that diction model in Example 1. Define G = (V, A ) to be the network
this game is played only once. Fig. 1 illustrates a feasible interdic- over which the problem is solved, with node set V and arc set A.
tion solution (the set of dashed arcs) to an instance of the shortest Let T be the node-arc incidence matrix having a row correspond-
path interdiction problem, where the follower takes the shortest ing to every node and a column corresponding to every arc. The
path from node 1 to node 6, and the leader’s budget allows it to column corresponding to arc (i, j) ∈ A has a 1 in row i, a −1 in row
interdict up to three arcs. The values in parentheses correspond j, and zeros elsewhere in the column. Let b ∈ Z|V | be the vector of
to interdiction effects. Prior to the interdiction, the shortest path all zeros except for a 1 in the row corresponding to the origin and
taken by the follower is 1 → 3 → 2 → 4 → 6, with a value of 17; af- a −1 in the row corresponding to the destination, where Z|V | is
ter the interdiction, the shortest path becomes 1 → 3 → 2 → 5 → 6, the set of all |V|-dimensional integer vectors. It is well-known that
with a value of 21.  there exists an optimal integer solution, which corresponds to an
origin-destination path, to any linear program constrained (only)
Besides the shortest path interdiction problem described above, by T y = b and y ≥ 0, if the objective coefficients on the y-variables
there are many other variants of basic interdiction problems where are nonnegative (or, more generally, if no negative-cost cycle ex-
assumptions A1–A5 are satisfied, including: ists). Let ci j ∈ R+ be the nominal cost of using arc (i, j) ∈ A, and
di j ∈ R+ be the amount by which arc (i, j) is lengthened if it is in-
(i) The binary knapsack interdiction problem (Caprara, terdicted, where R+ denotes the set of nonnegative real numbers.
Carvalho, Lodi, & Woeginger, 2016; Fischetti, Ljubic, Monaci, That is, the cost of using arc (i, j) is ci j + di j xi j , where xi j = 1 if arc
& Sinnl, 2019), where the leader and the follower each (i, j) is interdicted, and xi j = 0 otherwise. We express:
pick items according to their own knapsack constraints, but   
instead of maximizing its own profit, the leader minimizes (x ) = min ci j + di j xi j yi j (4a)
the follower’s maximum profit by blocking items that the (i, j )∈A
follower can choose; s.t. Ty = b (4b)
(ii) The clique interdiction problem (Mahdavi Pajouh, Boginski,
& Pasiliao, 2014), where the leader selects a set of edges up y ≥ 0. (4c)
to a given cardinality limit to remove from an undirected Note in this example that (4) is a simple shortest path problem
graph so that the size of a maximum clique in the remaining when x is fixed, and that Y(x) does not actually depend on x. To
graph is minimized; finish this model, we need only to represent the set X. Typically,
(iii) The matching interdiction problem (Zenklusen, 2010), where interdiction decisions are “all-or-nothing” (binary) decisions, and
the leader removes a limited subset of vertices or edges in can be constrained by a single cardinality constraint, by a knapsack
an undirected network under a given budget, such that the constraint, or via a more general set of constraints. If, for instance,
weight of a maximum matching in the remaining graph is the leader were restricted to interdict no more than γ arcs, where
minimized; γ ∈ Z+ , then
(iv) The project interdiction problem (Brown, Carlyle, Harney,  
Skroch, & Wood, 2009), where the interdictor attempts to 
maximally delay the completion time of a project by allo- X= x ∈ {0, 1}|A| : xi j ≤ γ . (5)
cating resources to disrupt its process. (i, j )∈A
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 799

Network interdiction problems (1) with (x) defined in (4) are that incorporates information asymmetry, as we will describe in
NP-hard in general. Ball, Golden, and Vohra (1989) have shown Section 4.4.
that even for the special case when X is defined as (5), and dij
is arbitrarily large for all (i, j) ∈ A (effectively removing arc (i, j) 1.3.2. Robust optimization models
from the network), the problem remains NP-hard. (This problem In a robust (maximization) optimization model, the leader
is known as the k-most-vital-arcs problem.) makes decisions under uncertain information. This uncertainty can
affect the leader’s objective and feasible region. The standard as-
Remark 1. The interdiction problem given in (1) maximizes the
sumption in robust optimization is that uncertainty outcomes oc-
interdiction effect subject to an interdiction cost budget. From a
cur in a way that minimizes the leader’s objective. Furthermore,
strategic planning perspective, where one pursues the best tradeoff
when uncertainty affects the feasible region, uncertainty reveals
between maximizing the interdiction effect and minimizing inter-
itself in a way that will make the leader’s decisions infeasible if
diction cost by finding the Pareto-optimal solutions associated with
possible. For brevity in this discussion, however, we assume that
the two objectives, a multi-objective network interdiction model
uncertainty is confined to the leader’s objective. Also, we assume
can be formulated and solved (Rocco & Ramirez-Marquez, 2010;
that the objective function is linear. The problem can be written in
Rocco, Ramirez-Marquez, & Salazar, 2010; Royset & Wood, 2007).
the form
max c x (8a)
1.3. Relationship to bilevel and robust optimization
s.t. x ∈ X. (8b)
Two areas closely related to interdiction models are bilevel op-
Here, we assume that c ∈ Rn
is unknown and that X is known.
timization and robust optimization. For completeness, we discuss
(See, e.g., Ben-Tal, El Ghaoui, & Nemirovski, 2009 for a more gen-
those modeling strategies here in order to contrast them with in-
eral analysis of robust equivalent formulations in which constraints
terdiction studies.
that describe the set X have uncertain data.) Although c is not
known precisely, the robust optimization model assumes that it
1.3.1. Bilevel optimization models belongs to an uncertainty set, C ⊆ Rn . Furthermore, given any de-
Bilevel optimization models regard a leader and a follower play- cision xˆ selected by the leader, we assume in robust optimization
ing a two-stage game just as in the interdiction model. However, that c takes on the value of a vector in C that minimizes c xˆ .
these models are more general than interdiction models in the fol- In practice, C often represents a polyhedral set of potential vec-
lowing ways: (i) the leader’s objective does not necessarily max- tors for c. For instance, let c be a nominal objective vector, per-
imize the minimum objective obtainable by the follower, and (ii) haps representing the anticipated objective coefficients. The bud-
the follower’s actions may now affect the leader’s objective and geted uncertainty set of Bertsimas and Sim (2004) assumes that c
feasible region. Hence, we denote the leader’s objective function as belongs to a hyperrectangle (simple lower and upper bounds on
f L and the follower’s objective function as f F . Because the leader’s each element of the vector), along with a side constraint that re-
feasible region now depends on the follower’s decisions, it is de- stricts c to be sufficiently close to c. The distance between c and
i=1 wi |ci − c̄i | for some
n
noted as X(y). The interdiction model given by (1) and (2) be- c is defined by a weighted 1-norm, e.g.,
comes: n
w ∈ R+ , which can typically be represented as a single additional
max f L (x, y ) (6a) constraint. More generally, we can define

s.t. x ∈ X (y ) (6b) c = c − y, (9)

y is an optimal solution to the follower’s problem, (6c) where y represents the possible deviation from c, and is con-
strained to belong to some set Y := {y ∈ Rn | ni=1 wi |yi | ≤  } for
where the follower’s problem is given as follows, with x being a some  > 0. The robust optimization model becomes:
fixed vector:  
min f (x, y )
F
(7a) max min (c − y ) x 
. (10)
x∈X y∈Y

s.t. y ∈ Y ( x ). (7b) Note that this model takes the form of (3), except that a lin-
Note that the choices of maximization and minimization for the ear objective is assumed in (10), and the set of possible objective
leader and follower, respectively, are arbitrary. The leader chooses coefficient deviations given by Y is not a function of the leader’s
x knowing that the follower will choose a y that optimizes prob- variables.
lem (7). The leader anticipates the follower’s behavior in (6c), seek- Although the basic interdiction setting considered has a simi-
ing to maximize the objective function (6a) and remain feasible to lar form to robust optimization problems with uncertain objective
(6b) based on knowledge of the follower’s behavior. Note that be- coefficients, they model different situations in practice. In network
cause (6c) requires the existence of a solution y that optimizes the interdiction problems, the agent that solves the network optimiza-
follower’s problem, the leader is forced to choose an x such that tion problem is the follower, with the leader acting first to inhibit
(7) has such an optimal solution. components of that network (such as lengthening the arc costs in
For the special case in which f L and f F are the same function, Example 1). In robust optimization problems, the network operator
and in which X(y) does not change based on y, we recover exactly acts first as the leader, and the follower simply degrades system
the interdiction model (2). However, the more general bilevel op- components based on the leader’s action to worsen the leader’s
timization model can capture the situation in which agents simply objective. Thus, in a robust shortest path problem, the leader com-
optimize based on their own self-interests. Dempe, Kalashnikov, mits to a path first and then the follower lengthens a worst-case
Pérez-Valdés, and Kalashnykova (2015) provide a comprehensive set of arcs with knowledge of the leader’s path choice.
survey on bilevel optimization. See DeNegre and Ralphs (2009), We refer the reader to the book of Ben-Tal et al. (2009), tuto-
Lozano and Smith (2017), Mitsos (2010), Saharidis and Ierapetri- rial chapters by Bertsimas and Thiele (2006) and Delage and Iancu
tou (2009), and Xu and Wang (2014) for a sampling of contempo- (2015), and an introductory article on basic robust optimization
rary algorithmic approaches in this area. Bilevel optimization mod- principles in Smith and Ahmed (2011). The relationship between
els facilitate one of the recent advances in network interdiction interdiction problems and robust optimization problems has been
800 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

utilized in the dynamic network interdiction setting, one of the

recent advances in network interdiction that we will describe in
Section 4.3.

1.4. Paper overview

The remainder of this paper is organized as follows.

Section 2 presents a brief history of network interdiction studies,
providing context for discussing recent studies in the interdiction
field. Section 3 examines major categories of algorithms that are Fig. 2. An illustration of a topological dual graph.
used to solve network interdiction problems. Section 4 explores
several types of contemporary and emerging studies on network
interdiction problems, showing how they differ from the as- no edges in its interior. We first add an artificial edge from the ori-
sumptions stated earlier in this section. Section 5 then discusses gin to the destination in G, and then initialize the topological dual
interdiction models for modern and future applications. by placing one dual node in every face of G. An (undirected) edge
connects two nodes in the dual graph if the nodes’ corresponding
faces are separated by a single edge in the primal graph, not count-
2. Origins of network interdiction models and algorithms
ing the artificial edge. One dual node located on the exterior of the
primal graph is the origin, and one dual node belonging to the face
Although the focus of this paper is on recent and emerging in-
bounded by the artificial edge is the destination. Fig. 2 illustrates
terdiction applications and algorithms, we first cover some of the
a primal graph with (lettered and shaded) dual nodes assigned to
classic network interdiction literature in this section.
each face. The corresponding topological dual network is shown on
One early example of a study on network flow problems having
the right side of this figure.
strong implications for the interdiction field stems from the max-
For each dual edge e, define (ie , je ) as the primal edge inter-
flow min-cut theorem established in the 1950s (Elias, Feinstein, &
sected by e. The traversal of dual edge e corresponds to placing
Shannon, 1956; Ford & Fulkerson, 1956). This theorem considers
primal edge (ie , je ) in the minimum cut for the primal. There-
a network G(V, A) with a source node s and destination node t,
fore, every dual origin-destination path corresponds to a primal cut
where each arc (i, j) ∈ A has a (positive, integer) capacity uij . The
set. For instance, the dual path A–C–E–G in Fig. 2 corresponds to
maximum flow value over G is the largest amount of flow that
cut set {(2, 4), (3, 4), (3, 5)}. The capacity of each edge placed in
can be sent from node s through the network such that flow is
the cut depends on whether or not it has been interdicted: pri-
conserved at all nodes i ∈ Vࢨ{s, t}, all flow emanating from s ter-
mal edge (i, j) has capacity 0 if interdicted, and uij if not. Thus,
minates at t, and the amount of flow on each arc (i, j) ∈ A is not
Wollmer’s approach solves the minimum cut problem by solving a
negative and not more than uij . A cut in G is a set of arcs A such
shortest path problem in the dual network, where the cost of using
that there exists no directed path from s to t in G(V, AࢨA ); the ca-
edge e depends on whether or not (ie , je ) is interdicted.
pacity of that cut is given by (i, j )∈A ui j . A minimum cut is a cut
Wollmer’s method creates an expanded dual network GE (VE , AE )
having the smallest capacity. With this notation thus defined, the
comprised of  + 1 copies of the dual network, where  is the num-
max-flow min-cut theorem states that the maximum flow value in
ber of edges that the leader can interdict. For each node i in the
G equals the capacity of a minimum cut in G. One particularly ele-
dual network, we place corresponding nodes i0 , . . . , i in VE . A path
gant proof of this theorem (Dantzig & Fulkerson, 1955) shows that
on the dual graph reaching ik , the kth copy of dual node i, indicates
the dual problem of the maximum flow linear programming for-
that k edges in that path correspond to interdicted edges. For each
mulation is exactly the minimum cut formulation. Indeed, as we
dual edge e we place edges (ike , jek ) in AE for all k = 0, . . . ,  and
will see later in this article, duality is one of the most important
(ike , jek+1 ) in AE for all k = 0, . . . ,  − 1. Edge (ike , jek ) corresponds to
tools used within interdiction studies.
an uninterdicted primal edge, and so its cost is given by cik jk = ue .
The implication of this theorem is that if the interdiction goal e e

is to disconnect two terminals of a network by removing arcs, and Edge (ike , jek+1 ) corresponds to an interdicted primal edge, and so
the effort required to remove a set of arcs is equal to the sum of its cost is given by cik jk+1 = 0. The cost of the shortest path in GE
e e
their arc capacities, then the interdiction problem can be cast as a from copy zero of the source node to copy k of the destination
min-cut problem (and thus solved as a max-flow problem). Harris node corresponds to the minimum cut value in the primal graph if
and Ross (1955) exploit this equivalence and apply it in the inter- the leader interdicts exactly k edges; thus, taking the minimum of
diction context for disconnecting rail networks. these values over all k at the destination node in GE corresponds
The more general maximum flow interdiction problem regards to the optimal objective value of this problem.
an interdictor who seeks to minimize the maximum flow that Note that planarity allows the existence of the topological dual,
can occur across the network, by removing a set of arcs from and the polynomiality of this procedure depends on the fact that
the network according to some budget constraints. Wollmer only  + 1 copies of the dual network are required in GE (recalling
(1964) examines the case in which G is planar, and an interdictor that  ≤ |A|). If a more general knapsack budget were used on a pla-
can remove up to a limited number  of arcs from the network. nar graph, i.e., interdicting primal arc (i, j) costs some positive inte-
These two assumptions (planarity and a cardinality-constrained ger ri j , and the total budget for interdiction is given by b, then an
interdiction budget) are essential to devising a polynomial-time adaptation of this procedure would become pseudopolynomial, be-
algorithm for the problem. cause we would require b + 1 copies of the planar dual network in
Wollmer’s approach utilizes a topological dual graph. For sim- GE . In fact, McMasters and Mustin (1970) examine exactly this case
plicity in defining this dual, we assume that the maximum flow for the problem of inhibiting the capability of a combatant force’s
problem takes place on a bidirected network, in which (i, j) ∈ A if resupply operations. Their algorithm relies on the generation of
and only if (j, i) ∈ A as well, where ui j = u ji . The results hold for di- minimal cuts and runs in exponential time. Phillips (1993) proves
rected networks as well with modifications; see Lawler (1976) for that the maximum flow interdiction problem is NP-hard (though
a discussion of topological duals. After embedding G in the plane, of course not in the strong sense) under a knapsack budget con-
define a face of G as a maximal subset of the plane that contains straint. If G were not planar, then the topological dual graph could
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 801

not be constructed at all. For this general case, Ghare, Montgomery, any fixed interdiction decision x. The simple idea of the “dualize-
and Turner (1971) and Ratliff, Sicilia, and Lubore (1975) develop and-combine” approach is to first take the dual formulation of the
combinatorial (and worst-case exponential) schemes for the prob- follower’s problem given a fixed x, so that both players’ problems
lem. Wood (1993) proves that the (cardinality-constrained) inter- share the same direction of optimization, and then release x as
diction problem is indeed strongly NP-hard for non-planar net- a decision vector, making it a single-level optimization problem.
works. Altner, Ergun, and Uhan (2010) provide improved bounds Specifically, let vector π be the dual vector associated with con-
on the approximability of this class of interdiction problems. straints (4b). The dual formulation for (4) can be written as:
Fulkerson and Harding (1977) consider a shortest path interdic-
(x ) = max b π (11a)
tion model in which each arc can be lengthened by any continu-
ous amount. The effort required to lengthen an arc is given by a
linear function of the amount by which the arc is lengthened, and
s.t. T  π ≤ c + Dx , (11b)
the leader is assumed to have a budget that constrains total inter-
 
where D = diag {di j : (i, j ) ∈ A} . Combining (11) with the leader’s
diction effort. The problem can be solved using parametric linear problem (1) gives the following single-level problem:
programming analysis, and Fulkerson and Harding (1977) show the
equivalence of this problem to solving a minimum cost network max b  π | T  π ≤ c + Dx , x ∈ X , (12)
flow problem, which is polynomial-time solvable. An alternative which can be readily solved via a standard optimization solver. We
characterization of this problem, which can also be solved by min- remark that problem (12) belongs to a general class of optimization
imum cost network flows, appears in Golden (1978). This problem problems called “reverse optimization problems” (see, e.g., Nguyen,
seeks to minimize the total interdiction effort required to increase 2016). For a given optimization problem, the corresponding reverse
the follower’s shortest path to a certain threshold value. However, problem is to modify some parameters under a budget constraint
when the interdiction decisions are binary, Ball et al. (1989) show so that the optimal value is improved as much as possible.
that the shortest path interdiction problem is NP-hard. When interdiction affects the feasible region of the follower’s
These studies led to the seminal work of Wood (1993), who optimization problem (according to Y(x)) instead of the objec-
provides a general mathematical programming modeling technique tive function, one could still apply the “dualize-and-combine” ap-
for network interdiction problems. The contributions are general proach, but with a little more effort. This situation typically arises
enough to handle situations in which interdiction actions can be in network flow interdiction problems in which interdiction deci-
discrete or continuous, and in which there may exist multiple in- sions reduce or eliminate arc capacities. For example, consider the
terdiction resources and flow commodities. Wood (1993) also ac- formulation:
counts for networks having directed or undirected edges, and for
those having multiple sources and destinations. (x ) = min c y (13a)
Early applications of network interdiction problems focused on
defense strategies, as described above. Currently, interdiction is s.t. Ty = b (π ) (13b)
applied in many different settings to model situations in which
one wishes to identify vulnerabilities to accidental or intentional yi j ≤ ui j ( 1 − xi j ), ∀(i, j ) ∈ A (−βi j ) (13c)
disruptions to infrastructure. In many other studies, the interdictor
is taken to be the protagonist in the problem scenario, such as yi j ≥ 0, ∀(i, j ) ∈ A, (13d)
in the study of Washburn and Wood (1995) where the interdictor
seeks to detect evaders (see also Morton, Pan, & Saeger, 2007; where parameters uij are simple upper bounds on yij (e.g., all ui j =
Sullivan, Morton, Pan, & Smith, 2014a for studies in which the 1 for the shortest path interdiction problem). Hence, (13c) enforce
interdictor’s aim is to thwart nuclear smuggling efforts). As an- the condition that yij ≤ uij unless arc (i, j) is interdicted, in which
other example, hospital infection containment is considered by case yij is forced to equal zero.
Assimakopoulos (1987), where interdiction actions seek to limit Following the “dualize-and-combine” approach directly, we as-
the spread of infection. sociate constraints (13b) with dual vector π and each constraint in
(13c) with a dual variable −βi j . The resulting single-level optimiza-
tion problem becomes:
3. Common algorithmic approaches for network interdiction 
max b π − ui j βi j (1 − xi j ) (14a)
In this section, we survey various algorithms in the network in- (i, j )∈A
terdiction literature that rely on different assumptions made on the
s.t. T  π − I|A| β ≤ c (14b)
interdiction problems, using the basic interdiction setup presented
in Section 1.1. In Section 3.1 we examine the case in which the
follower’s problem (2) can be modeled as a convex optimization βi j ≥ 0, ∀(i, j ) ∈ A (14c)
problem, and then describe the more complicated case when the
follower’s problem is nonconvex in Section 3.2. x ∈ X, (14d)
where I|A| is the |A| × |A| identity matrix. Although a complicating
3.1. Network interdiction with convex follower’s problem bilinear term βi j (1 − xi j ) shows up in the objective function (14a),
when the interdiction decision xij is restricted to be a binary vari-
We start by reviewing solution approaches for network inter- able (which is typically true for network interdiction problems),
diction problems where the follower’s problem can be formulated we could linearize the bilinear term using standard McCormick in-
as a convex optimization problem. The convexity assumption en- equalities given upper bounds Mij on β ij . This linearization strategy
ables duality-based approaches such as dualize-and-combine and would (i) add a new variable wi j to replace βi j (1 − xi j ) in (14a),
Benders decomposition, as we examine here. and (ii) add the following inequalities to (14):
wi j ≥ βi j − Mi j xi j , and wi j ≥ 0, ∀(i, j ) ∈ A.
3.1.1. Dualize-and-combine
For simplicity, we consider the shortest path interdiction prob- The typical linearization inequalities wi j ≤ βi j and wi j ≤ Mi j (1 −
lem where the inner problem is a shortest path problem (4) given xi j ) are not necessary here because wi j will take the smallest value
802 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

possible at optimality, and these upper bounds on wi j can thus be Ad-hoc formulation example. In the “dualize-and-combine” ap-
omitted from the formulation. proach, optimality of the follower’s decision is enforced by the dual
An alternative reformulation of (13) is to penalize the use of formulation via strong duality. Depending on the problem struc-
an interdicted arc rather than by prohibiting the use of an inter- ture, an alternative characterization of the follower’s optimal de-
dicted arc within the constraint set. The validity of this reformu- cisions can be utilized to construct ad-hoc formulations. A classic
lation again relies on the binary restriction on decision variables example of these ad-hoc formulations is the one proposed for the
xij . For all (i, j) ∈ A, given an upper bound Mij on the optimal dual uncapacitated facility interdiction problems by Church, Scaparra,
multiplier β ij associated with (13c) for all feasible x ∈ X, formula- and Middleton (2004). This problem examines a set of facilities F
tion (13) can be reformulated as: that is under attack from an interdictor that can essentially elim-
inate some r ∈ Z+ facilities (0 < r < |F|). There exists a set N of
(x ) = min c y + x My (15a) demand points, and the follower assigns each point to the clos-
est facility that is not attacked, so that the total routing cost is
s.t. Ty = b (15b) minimized. Letting dij denote the distance from demand point i to
facility j, the facility interdiction problem can be formulated as the
following max-min problem:
∀(i, j ) ∈ A,
0 ≤ yi j ≤ ui j , (15c) 
  max min di j yi j (16a)
where M = diag {Mi j : (i, j ) ∈ A} . One can then apply “dualize- x∈X
i∈N j∈F
and-combine” for formulation (15). 
We now show that this reformulation is valid, i.e., formu-
s.t. yi j = 1, ∀i ∈ N (16b)
j∈F
lation (15) is equivalent to formulation (13). Given a fixed
x∗ ∈ {0, 1}|A| , let z1 (x∗ ) and z2 (x∗ ) be the optimal objective value yi j ≤ 1 − x j , ∀i ∈ N, j ∈ F (16c)
of (13) and (15), respectively. We show that z1 (x∗ ) = z2 (x∗ ). First,
let y∗ be an optimal solution to (13) with x = x∗ . Observe that yi j ∈ {0, 1}, ∀i ∈ N, j ∈ F , (16d)
Eq. (13c) implies that x∗ My∗ = 0, so y∗ is feasible to (15) with an
where X = {x ∈ {0, 1}|F | | j∈F x j ≤ r}. The constraint matrix of
objective value of z1 (x∗ ) = c y∗ , thus z1 (x∗ ) ≥ z2 (x∗ ). On the other
the inner problem for (16) is totally unimodular, implying that
hand, let ȳ∗ be an optimal solution to (15) with x = x∗ . We claim
the integrality restrictions on yij variables can be relaxed without
that ȳ∗i j = 0 whenever x∗i j = 1. This is because the dual price for
changing the optimal objective value of the inner problem. Thus
constraint (13c), which can be interpreted as the profit of having the follower’s problem can be treated as a linear program. Instead
one extra unit of resource yij , is upper bounded by the cost Mij . of dualizing this inner problem, Church et al. (2004) control both
Therefore, z2 (x∗ ) = c ȳ∗ , and ȳ∗ is feasible to (13) with x = x∗ , im- the leader’s decision x and the follower’s decision y in a single-
plying that z2 (x∗ ) ≥ z1 (x∗ ). This idea was first proposed by Morton level formulation, and explicitly enforce optimality conditions for
and Wood (1999), and has been a popular reformulation tool for the follower’s decision y as follows. Given any xˆ ∈ X, there exists
solving maximum flow interdiction problems (Cormican, Morton, an optimal y∗ so that for each demand point i ∈ N, there is only
& Wood, 1998). one index ji∗ that y∗i j∗ = 1 and y∗i j = 0 for all j = ji∗ . This index ji∗
In both of the above approaches, the upper bound Mij for β ij i

variables is crucial. In practice, it is important to find a relatively is one belonging to the set:
tight bound Mij on β ij , in order to obtain a strong linear pro- arg min j∈F (xˆ ) {di j },
gramming relaxation bound. These bounds can be obtained by
where F (xˆ ) := {k ∈ F : xˆk = 0}. Define Tij := {k ∈ F|dik > dij } as the
gleaning insights from the follower’s problem structure. For exam-
set of facilities that are farther from point i than facility j. Together
ple, Cormican et al. (1998) provide such upper bounds for maxi-
with (16b), logical constraints x j = 0 ⇒ yik = 0 for all k ∈ Tij ensures
mum flow interdiction problems, and Lim and Smith (2007) derive
that y is optimal to the follower’s optimization problem. Using lin-
dual upper bounds for multi-commodity network flow interdiction
ear constraints (17d) below to model these logical constraints, the
problems.
uncapacitated facility interdiction problem can be reformulated by
the following single-level integer programming problem:
Continuous interdiction. When the interdiction decisions x are 
modeled as continuous variables, neither the standard lineariza-
max di j yi j (17a)
i∈N j∈F
tion approach nor the penalty-based approach described above can 
be applied when the interdiction affects the feasible region of the s.t. xj = r (17b)
follower’s problem. (When interdiction only affects the follower’s j∈F

objective, the duality-based reformulation (12) is still valid.) Rec- yi j = 1, ∀i ∈ N (17c)
ognizing that variables involved in the bilinear terms are disjointly j∈F
constrained, specialized finitely convergent algorithms proposed in, 
e.g., Sherali and Shetty (1980) or Alarie, Audet, Jaumard, and Savard
yik ≤ x j , ∀i ∈ N, j ∈ F (17d)
k∈Ti j
(2001) can be applied, although these algorithms may suffer from
computational and implementation challenges. When the leader’s x j ∈ {0, 1}, ∀j ∈ F (17e)
feasible set X is a polyhedron characterized as a single-row bud-
get constraint, Lim and Smith (2007) take advantage of the ex- yi j ∈ {0, 1}, ∀i ∈ N, j ∈ F. (17f)
treme point solution structure of this polyhedron, noting that all
but one of the variables are nonbasic. They propose a partitioning 3.1.2. Benders decomposition
algorithm by designating one variable at a time as the basic vari- Benders decomposition is an outer-approximation technique for
able, while all others are fixed at either of their bounds, i.e., 0 or the hypograph defined by {(x, θ )|θ ≤ (x)}, which is a convex
1, which can therefore be modeled as binary variables. Standard closed set when the value function (x) is concave. Benders de-
linearization approaches such as McCormick inequalities can then composition starts from a reformulation of the follower’s problem
be applied to reformulate the problem as a mixed integer linear defined in terms of all its extreme point solutions. Taking again the
program. shortest path interdiction problem as an example, let Y denote the
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 803

feasible region of the follower’s problem (4), and let Yˆ be the set interdiction problem that motivates the idea of Salmerón et al.
of all extreme points of Y. Then the leader’s problem (1) can be (2009), where (x) is a multilinear nonconvex function. The key
reformulated as: is to find a valid “penalty vector” v(xˆ ) so that the following global
Benders cut is valid for all x ∈ X:
max θ (18a)
   θ ≤ (xˆ ) + v(xˆ ) (x − xˆ ). (21)
s.t. θ≤ ci j + di j xi j yˆi j , ∀yˆ ∈ Yˆ (18b) When the value function (x) is concave, and standard Ben-
(i, j )∈A ders decomposition can be applied, vector v(xˆ ) corresponds to the
x ∈ X. (18c) gradient or subgradient of (x), obtained by solving the follower’s
problem with a fixed xˆ . However, when (x) is not concave, in
There are typically exponentially many extreme point solutions
order to construct a valid inequality (21) with a valid coefficient
contained in the set Yˆ , preventing a direct use of formulation (18).
vector v(xˆ ), one must exploit specific problem structures such as
In Benders decomposition, a master problem that enforces con-
the binary property of the interdiction decisions.
straints of the form (18b) for only a subset of the extreme points
Salmerón et al. (2009) apply the idea of global Benders decom-
Y ⊆ Yˆ is constructed. This master problem is therefore a relaxation
position for interdiction problems in electric power grid networks,
of (18). Given a solution xˆ to the master problem, the following
where interdiction is interpreted as a way to perform a worst-
subproblem is solved, and the corresponding optimal solution yˆ is
case vulnerability analysis for a power grid. The primary goal of an
then added to the subset Y :
  imaginary interdictor is to maximize the minimum power genera-
 tion cost and load shedding. In this case, the value of v(xˆ ) for each
min (ci j + di j xˆi j )yi j | y ∈ Y . (19) component i in the power grid can be chosen by finding the max-
(i, j )∈A imum possible load that can be recovered from “un-interdicting”
The master problem (18) with Y and the subproblem (19) are component i, i.e., changing xˆi = 1 to xˆi = 0, and the maximum pos-
iteratively solved in an alternating fashion until convergence. sible load shedding from interdicting component i, i.e., changing
The convergence of Benders decomposition is guaranteed by the xˆi = 0 to xˆi = 1. Power generation capacity on each component i
finiteness of Yˆ . Interpreted as a cutting plane approach, Benders helps to identify these bounds.
decomposition can be potentially accelerated via various stabi-
lization approaches such as the level bundle method (Lemaréchal, 3.2. Network interdiction with a nonconvex follower’s problem
Nemirovski, & Nesterov, 1995), cut selection strategies (Magnanti
& Wong, 1981; Smith, Lim, & Alptekinoglu, 2009), etc., and can In this section, we relax the convexity assumption on the fol-
be integrated within the branch-and-bound algorithm for solving lower’s problem, which is crucial for the “dualize-and-combine”
(18) when X involves integer restrictions. approach introduced in Section 3.1 (and may also be needed for
the Benders decomposition approach). Network interdiction prob-
Enhancements of Benders decomposition via combinatorial structures. lems involving a nonconvex follower’s problem remain challenging
When the leader’s feasible set X involves integrality restrictions, to solve; however, much progress has recently been made on im-
one could further strengthen the formulation by generating in- portant special cases of these problems.
equalities that cut off not only (fractional) relaxation solutions
but also integer feasible solutions, so long as at least one optimal Defender–attacker–defender problems. One important class of prob-
solution has either been recorded by the algorithm or is not made lems in which the follower’s problem is nonconvex is the
infeasible by the cut. These cuts are referred to as super-valid defender–attacker–defender model (DAD), see, e.g., Brown, Car-
inequalities by Israeli and Wood (2002). Given an integer feasible lyle, Salmerón, and Wood (2006). This model extends the clas-
solution xˆ , a cut corresponding to yˆ is generated by solving sical leader–follower paradigm in network interdiction, which is
subproblem (19), and the following super-valid inequality is added also known as an attacker–defender (AD) model (if one considers
to the Benders master problem: the interdicting agent as the “attacker”). In DAD models the de-
 fender makes a preliminary set of “fortification” decisions before
xa ≥ 1, (20) the attacker–defender game takes place, resulting in a three-stage
a∈A(yˆ ) game. The defender’s fortification decisions typically serve to pro-
tect infrastructure against potential interdictions.
where A(yˆ ) := {a ∈ A | yˆa = 1}, i.e., the set of arcs corresponding
For instance, in the shortest path interdiction example formu-
to the shortest path chosen by the follower in response to the
lated in (14), a DAD model may allow the defender to first select
leader’s decision xˆ . The underlying idea behind (20) is that given
some δ ∈ Z+ arcs to fortify, where a fortified arc cannot be inter-
an incumbent interdiction decision, xˆ , at least one arc on the
dicted. The DAD model can be formulated as the following two-
shortest path corresponding to xˆ must be interdicted in order to
stage min-max model:
potentially increase the follower’s shortest path value. Super-valid 
inequalities (20) can be further strengthened based on the best min max b π − ui j βi j (1 − xi j ) (22a)
w∈W
upper bound obtained so far. Substantial computational time (i, j )∈A
reduction from utilizing super-valid inequalities in Benders de-
composition was observed by Israeli and Wood (2002) for the
s.t. T  π − I|A| β ≤ c (22b)
shortest path interdiction problem.
βi j ≥ 0, ∀(i, j ) ∈ A (22c)
Global Benders decomposition. Salmerón, Wood, and Baldick
xi j ≤ 1 − wi j , ∀(i, j ) ∈ A (22d)
(2009) extend the scope of Benders decomposition to models
where the value function (x) of the interdiction decision x is not
x ∈ X, (22e)
necessarily concave. For example, this is the case for network flow
interdiction problems in which interdictions reduce or eliminate where W = {w ∈ {0, 1}|A|
| (i, j )∈A wi j ≤ δ}. Observe that
arc capacities, where the value function (x) is a convex function (22d) prohibits the interdiction of fortified arcs.
of x, because it can be written as the maximum of a set of affine Formulation (22) exhibits a property commonly found in DAD
functions. More complicated cases include the optimal power flow problems: The inner maximization problem is nonconvex, and
804 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

there exists no obvious way to obtain a strong dual of the in- subset of the follower’s solutions Yˆ ⊆ Y . Given such a subset Yˆ , the
ner problem to combine with the outer problem. Smith, Lim, optimal objective value of the following model gives the maximum
and Sudargho (2007) use a penalty-based reformulation similar perceived damage made by the leader:
to that used in (15a). In particular, there exists a penalty matrix
zP (Yˆ ) := max min f (x, y ), (26)
M = diag(Mi j )(i, j )∈A for appropriately large Mij -values such that x∈X y∈Yˆ (x )
(22) yields the same set of optimal solutions as does the follow-
ing formulation: where Yˆ (x ) = Yˆ ∩ Y (x ). Value zP (Yˆ ) is the objective function value
 anticipated by the leader, if the leader is blind to all other pos-
min max b π − ui j βi j (1 − xi j ) + w Mx (23a) sible solutions not in Yˆ that the follower can make. Letting xˆ be
w∈W
(i, j )∈A an optimal solution to (26), the real damage caused by decision xˆ
(evaluated by considering all possible follower’s decisions) is given
s.t. T  π − I|A| β ≤ c (23b)
by:

βi j ≥ 0, ∀(i, j ) ∈ A (23c) zR (xˆ ) := min f (xˆ , y ). (27)

y∈Y (xˆ )

Clearly, zR (xˆ ) and zP (Yˆ ) provide lower and upper bounds on the
x ∈ X. (23d)
optimal objective value z∗ respectively, i.e., zR (xˆ ) ≤ z∗ ≤ zP (Yˆ ). If
The penalty terms enforce the condition that wi j xi j = 0, instead the two bounds are not equal, then the corresponding optimal fol-
of relying on explicit constraints to serve that role. As such, the lower’s decision yˆ in (27) must not belong to Yˆ , and therefore we
feasible region of the inner problem in formulation (23) does not can update Yˆ by Yˆ := Yˆ ∪ {yˆ }. Additional samples of feasible solu-
depend on w, although the objective function still does. Given tions in Y can be generated to supplement Yˆ , while other elements
a fixed fortification w ˆ , problem (23) becomes a disjointly con- of Yˆ can be removed depending on their corresponding objective
strained mixed integer bilinear programming problem. Define function values.
as the set of all vectors (π , β) that satisfy constraints (23b) and The algorithm then solves (26) with the updated set Yˆ of fol-
(23c), and let ext( ) be the extreme points of this polyhedron. lower’s decisions. The finiteness of the algorithm is guaranteed by
Consider a fixed w. For a given xˆ , the problem is a linear program the finiteness of the leader’s feasible set X. The success of this ap-
in the remaining variables (π , β). Similarly, for a fixed (π ˆ , βˆ ), the proach relies on the following aspects: (i) the increasing size of the
problem is a mixed integer program in variables x. Therefore, an sampled solution set Yˆ should be well-maintained so that the per-
optimal solution (π ˆ , βˆ , xˆ ) to the inner problem of (23) exists in ceived damage problem (26) can be solved efficiently; (ii) an effi-
which (π ˆ , βˆ ) ∈ ext( ) and xˆ ∈ X. Because the inner problem is a cient sampling approach that supplies a diverse sample in Y needs
maximization problem, the optimal objective value to this prob- to be employed; and (iii) a reasonably efficient approach should
lem, given an interdiction vector w ˆ , is given by: be available to solve the real damage problem (27). Note that the
  backward sampling approach is similar to that of standard Ben-
 ders decomposition. A master problem with a restricted set Yˆ ⊆ Y
max b π

ˆ − ui j βˆi j (1 − xˆi j ) + w
ˆ  Mxˆ . (24)
xˆ ∈X, (π
ˆ ,βˆ )∈ext( ) of the follower’s decisions is solved to yield an upper bound, while
(i, j )∈A
a subproblem with a fixed leader’s decision xˆ is solved to yield a
This formulation allows us to convert (22) into a single-level for- feasible solution (which is used to update the lower bound).
mulation as follows, albeit with an exponential number of con- Moreover, this study also examines the DAD setting, using
straints. inequalities similar to (25b) to guide the fortification decisions.
min θ (25a) The given approach employs the bounds obtained by computing
real and perceived damage values given fortification decisions, and
 uses the objective function bound from the best obtained solu-
s.t. θ ≥ b πˆ − ui j βˆi j (1 − xˆi j ) + w Mxˆ , tion (the “incumbent value”) to reduce the number of backward
(i, j )∈A
sampling searches. In particular, when the backward sampling
for all xˆ ∈ X and (π
ˆ , βˆ ) ∈ ext( ) (25b) bounds indicate that a fortification decision cannot improve the
incumbent value, the backward sampling search corresponding to
w ∈ W. (25c) that fortification vector is terminated and a combinatorial Benders
inequality (Codato & Fischetti, 2006) is generated to cut off that
Benders decomposition can be applied by adding inequalities fortification solution. Other algorithmic enhancements for the DAD
(25b) only as needed during the solution process. This process algorithm are detailed in Lozano and Smith (2017).
would solve a relaxation of (25) with a small subset of constraints
(25b) present. Then, given an optimal solution w ˆ to that relax- Benders primal decomposition. When the follower’s problem is
ation, the inner problem to (23) is next solved to determine the ac- modeled as a mixed integer program with a finite number of in-
tual value of the attacker–defender problem given w ˆ . If this value teger feasible solutions, the Benders primal decomposition approach
matches the value of the relaxation objective obtained from solv- (also known as the column-and-constraint generation method, see,
ing (25), then the algorithm terminates with an optimal solution. e.g., Zeng & Zhao, 2013) can alternatively be employed. In contrast
Otherwise, an inequality of the form (25b) is generated and added to standard Benders decomposition, where the number of deci-
to the relaxation, and another iteration of the Benders decomposi- sion variables is fixed, Benders primal decomposition iteratively
tion algorithm is executed. tightens an approximation to the extended single-level reformu-
lation of the problem by adding both new decision variables and
Backward sampling approach. Lozano and Smith (2017) propose new constraints. These new decision variables correspond to the
a backward-sampling approach for solving interdiction problems follower’s response given a fixed leader’s decision xˆ , and the new
where the interdiction decisions x are binary variables, i.e., X ⊆ constraints enforce the optimality condition for the follower’s
{0, 1}nL . The idea of this approach is based largely on that of Ben- response. This approach has proved to be effective for applications
ders decomposition as described above. Instead of capturing all of network interdiction problems arising in vulnerability analysis
possible follower’s decisions via constraints using strong duality, of power systems, see, e.g., Zhao and Zeng (2013), Yuan, Zhao, and
this approach partially enumerates (via sampling) and maintains a Zeng (2014), Yuan et al. (2016), and Wu and Conejo (2017).
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 805

Iterative convex restrictions. When the follower’s problem is mod- stating the probability that the evader uses path h ∈ H. Also, let Q
eled as a mixed integer program, another possible solution ap- be an |A| × |H| matrix, with qah = pa if arc a belongs to path h and
proach is based on the creation of convex approximations to the qah = 0 otherwise, for all a ∈ A and h ∈ H. Given a solution (xˆ , yˆ ),
follower’s problem. This approach is taken by Tang, Richard, and the probability of detection is computed as xˆ T Q yˆ . Washburn and
Smith (2016) for interdiction problems in which the leader’s vari- Wood (1995) seek to find a Nash equilibrium solution (x∗ , y∗ ), i.e.,
ables are binary and the follower solves a mixed integer program. one in which
In their model, the leader’s interdiction decisions affect the fol-
lower’s feasible region but not the objective.
y∗ ∈ arg miny≥0:eT y=1 (x∗ ) Qy , and
The algorithm builds a convex inner-approximation to the x∗ ∈ arg maxx≥0:eT x=1 x Qy∗ ,
follower’s feasible region, which can be solved by “dualize-and-
combine” (see Section 3.1.1), yielding a lower bound on the overall where e is a vector of all ones (having conforming dimension)
problem. Much like the backward sampling approach described above. That is, the Nash equilibrium solution satisfies the condition
above, given the leader’s solution, the algorithm then solves the that neither the interdictor nor the evader wish to change their
actual follower’s problem as a mixed integer program and com- strategy, having observed the other’s strategy. Among the contribu-
pares the objective value that the follower would obtain given this tions in Washburn and Wood (1995) is a linear-programming based
interdiction with the lower bound previously computed. If these algorithm to solve this problem in which the evader encodes their
values do not match, then the convex inner approximation to the solution via a (polynomial) network flow solution as opposed to
follower’s problem is expanded accordingly and the algorithm re- requiring the enumeration of all (exponentially many) paths.
iterates until an optimal solution is provably obtained. The critical Goldberg (2017) considers a form of this game with three ma-
step in this procedure lies in the formulation of this convex inner jor differences. The first difference is that the Washburn and Wood
approximation: if the inner approximation is merely a convex hull (1995) model essentially treats the x-variables as the amount of ef-
of the follower’s solutions that have already been observed, then fort in monitoring the arcs, where the probability of detection on
the algorithm performs poorly. However, if the inner approxima- arc a is a linear function (pa xa ) of the interdiction effort applied
tion implicitly includes many follower solutions that have not to arc a. In the problem considered by Goldberg (2017), that prob-
been explicitly computed thus far, then the algorithm performs ability becomes a nonlinear function of xa , where the function is
far better. The latter case has shown to arise in binary knapsack assumed to be logarithmically convex and decreasing. The second
interdiction and clique interdiction problems according to Tang difference is a set of multiple possible origins and destinations that
et al. (2016). the evader may use. The third difference is that the evader’s pay-
off is a function, c, of the origin-destination pair selected, multi-
4. Recent and emerging interdiction studies plied by the total probability of evasion on the path selected (given
by the product of evasion probabilities on the path). The interdic-
We now discuss some recent interdiction studies, especially tor’s payoff is the negative of a possibly different function, d, of the
those that relax the assumptions made in Section 1 of this paper. evader’s origin-destination pair, multiplied by the total probabil-
Section 4.1 regards problems in which both agents act simulta- ity of evasion on the path. Because c and d might be distinct, the
neously, effectively modeling the situation when the follower is problem is no longer a zero-sum game. Goldberg (2017) shows that
unaware of the leader’s actions. Section 4.2 examines the problems Nash equilibrium solutions must still exist for this game, and can
where data or interdiction actions are uncertain. Section 4.3 dis- be computed (exactly or approximately depending on certain as-
cusses dynamic interdiction problems involving repeated interac- sumptions) in polynomial time.
tion between the agents within one game. Section 4.4 considers Another application arises in a game studied by Baykal-Gürsoy,
interdiction problems in which information asymmetry exists. Duan, Poor, and Garnaev (2014) between an adversary and a first
Section 4.5 explores problems in which information is incomplete responder. The simplest version of their game involves a set of
for either the leader or follower. nodes, each of which is associated with an occupancy value (e.g.,
the number of civilians at that location), and a detection prob-
4.1. Simultaneous play ability if monitored. The adversary’s payoff for attacking a node
is proportional to the node’s occupancy, multiplied by the prob-
The case in which the two agents play simultaneously gives ability that the first responder fails to detect the adversary at the
rise to a large field of research. While this research is largely be- node. This problem yields a zero-sum game in which the adver-
yond the scope of this paper, it is instructive to highlight some sary seeks to maximize its payoff by selecting a mixed strategy
core principles behind simultaneous play as they relate to inter- for node attacks, while the first responder seeks to minimize pay-
diction. We begin by examining the work of Washburn and Wood off by selecting its own mixed strategy for monitoring nodes. An-
(1995) regarding a two-player interdiction problem involving an other key contribution of Baykal-Gürsoy et al. (2014) comes in ex-
evader, who wishes to choose a path to avoid detection, and an in- amining population flows among the nodes, as well as patrolling
terdictor, who can monitor any single arc. If the interdictor chooses options among a set of nodes that the first responder can take.
to monitor arc (i, j) ∈ A, then the evader is detected with probabil- The authors employ a partially observable Markov decision process
ity pij if the evader’s path includes arc (i, j). In this zero-sum game, (POMDP) model to analyze this dynamic version of the game. We
the evader seeks to minimize its probability of detection, whereas also refer the reader to Guan, He, Zhuang, and Hora (2017), Powell
the interdictor seeks to maximize the probability of detecting the (2007) and Zhuang and Bier (2007) for a sampling of recent simul-
evader. However, in this setting the evader cannot see the interdic- taneous interdiction games having national security implications.
tor’s actions before committing to its path. Therefore, both agents
essentially commit to their strategies simultaneously. 4.2. Stochastic network interdiction problems
This situation results in mixed-strategy solutions for both
agents: the interdictor selects a probability that it will interdict Two of the strongest assumptions of the classical network
each of the arcs in the network, and the follower selects a proba- interdiction problem stated in the beginning of Section 1.1 are
bility for choosing each path in the network. Let H be the set of all assumptions A1 and A2, where it is assumed that all problem data
origin-destination paths. Define xij as a variable stating the proba- is known to both the leader and the follower, and the leader is
bility that the interdictor monitors arc (i, j) ∈ A, and yh as a variable certain of the interdiction effect on the follower’s problem. Here,
806 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

we extend the scope of network interdiction problems into the ship between the SAA problem (29) and the corresponding orig-
realm of optimization under uncertainty by considering the case inal stochastic program (28) has been extensively studied in the
when the network interdiction problem involves uncertain data stochastic programming literature (see, e.g., Shapiro et al., 2009),
(such as the cost and interdiction effect on each arc), and the and most SNIP studies focus on solving the SAA problems such as
leader has to make an interdiction decision before the realization (29).
of uncertainty. However, we do assume that the leader has some
knowledge about the underlying probability distribution of the Computational aspects for solving SNIPs. From a computational per-
random variables that are used to characterize the uncertainty, spective, the “dualize-and-combine” approach is less competitive
possibly in the form of a finite set of scenarios. We also assume in solving (29) than algorithms that enable decomposition by sce-
that the follower is able to make decisions by solving a deter- narios such as Benders decomposition, especially when the num-
ministic optimization problem after observing both the realization ber of scenarios is large. These decomposition algorithms (e.g.,
of random variables and the leader’s decision. We refer to this Janjarassuk & Linderoth, 2008, for stochastic maximum flow in-
setting as a stochastic network interdiction problem (SNIP) with terdiction problems with interdiction effect uncertainty) can be
a “wait-and-see” follower. In the next section, we discuss the implemented in a parallel computing framework, leveraging high-
case when both the leader and the follower need to make their performance computing resources. Note that since the feasible set
decisions before the realization of random variables, which we X of interdiction decisions usually involves integrality restrictions,
refer to as an SNIP with a “here-and-now” follower. it is computationally more efficient to integrate Benders decompo-
For example, consider the shortest path interdiction prob- sition into a branch-and-bound framework. This so-called branch-
lem (4) with uncertain arc costs and interdiction effects and-cut algorithm typically outperforms a naive cutting plane ap-
(c˜(ξ ), d˜(ξ )), where ξ ∈ is a random vector that follows a known proach that solves an integer program at each iteration.
joint probability distribution. Since the leader makes decisions un- This line of research has developed strong valid inequalities for
der uncertainty, their risk preference plays a key role in terms of the stochastic maximum reliability path interdiction problem. In
formulating the objective function of the problem. Most literature this problem, the follower seeks a path having the highest chance
on SNIPs assumes that the leader is a risk-neutral decision maker, of avoiding detection from an origin to a destination. The leader,
i.e., in the context of the shortest path interdiction, they maximize on the other hand, allocates interdiction budget to increase detec-
the expected shortest path value taken by the follower: tion rates on an optimally chosen subset of arcs so that this chance
is minimized. We note that the maximum reliable path interdic-
max E[ f (x, ξ )], (28)
x∈X tion problem is equivalent to the shortest path interdiction prob-
where f(x, ξ ) represents the shortest path value taken by the fol- lem if the chance of being detected on each arc is less than 1. The
lower under each realization ξ : leader’s knowledge of the follower’s origin-destination pair is as-
 sumed to be uncertain, and is characterized by a finite set of sce-
f (x, ξ ) := min ci j ( ξ ) + di j ( ξ )xi j yi j narios. Motivated by the mixing structure (Atamtürk, Nemhauser,
(i, j )∈A & Savelsbergh, 20 0 0; Günlük & Pochet, 2001) of the underlying
s.t. T y = b large-scale mixed integer programming formulation, Pan and Mor-
ton (2008) propose a set of strong valid inequalities called “step in-
yi j ≥ 0, ∀(i, j ) ∈ A.
equalities” to strengthen the linear programming relaxation bound
SNIPs such as (28) are a special case of stochastic programs of the formulation. Sullivan, Smith, and Morton (2014b) general-
(Shapiro, Dentcheva, & Ruszczynski, 2009). Stochastic programs are ize these inequalities for problems involving asymmetric data per-
challenging to solve, as one can see in (28) that even with a fixed ceptions (see Section 4.4 for more on such problems) and pro-
interdiction decision x, computing the objective involves high- vide a convex hull analysis. Bodur, Dash, Günlük, and Luedtke
dimensional integration. Among all approaches that have been (2016) strengthen the generated Benders cuts by exploiting the
studied in the stochastic programming literature, two classes of ap- integrality of the interdiction decisions. More recently, Towle and
proaches have been particularly popular for SNIPs: (i) sequential Luedtke (2018) reformulate this problem using a path-based for-
approximation and (ii) sample average approximation. mulation, and propose additional valid inequalities exploiting the
Cormican et al. (1998) propose a sequential approximation ap- supermodularity structure (see, e.g., Nemhauser & Wolsey, 1988,
proach to solve SNIPs where interdictions may fail randomly (they Chapter III.3) resulting from the reformulation.
focus on SNIPs in the context of the maximum flow interdiction In addition to the risk-neutral model, where the leader max-
problems). These authors apply classical bounding techniques in imizes the expected interdiction effect, SNIPs with a risk-averse
stochastic programming (see, e.g., Kall & Wallace, 1994) on (28), leader have also been studied in the literature. Instead of opti-
and recursively tighten the bounds until the optimality gap is small mizing average performance over the long run, a risk-averse leader
enough by refining partitions of the support of random vector ξ . hedges against the risk of disastrous outcomes under certain risk
On the other hand, the sample average approximation (SAA) ap- measures such as the conditional value at risk, which typically
proximates the original distribution of ξ by its empirical distribu- fits the mindset of a defender, e.g., in a homeland security set-
tion (possibly via Monte Carlo sampling) in the form of a set of ting. For example, Atamtürk, Deck, and Jeon (2019) model the risk
scenarios {(ck , dk )}k∈N , where each scenario happens with proba- aversion of the leader by a mean-risk model and reformulate it
bility pk ( pk = 1/|N | if Monte Carlo sampling is used). using a convex quadratic mixed integer program. Song and Shen
  (2016) propose an alternative risk-averse model using chance con-
max pk min (cikj + dikj xi j )yki j (29a) straints (Shapiro et al., 2009). These works exploit not only the
x∈X yk
k∈N (i, j )∈A
combinatorial structures from the SNIPs but also additional struc-
s.t. T yk = b (29b) tures embedded within the risk measures.

yki j ≥ 0, ∀(i, j ) ∈ A. (29c) 4.3. Dynamic network interdiction

We note that there is one copy of the follower’s decision yk In this section, we focus on two zero-sum games. In both of
for each scenario k ∈ N , because the follower is given the flexibil- these games, players have full information about the network and
ity to choose different paths for different scenarios. The relation- its data. However, these games depart from assumption A4, under
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 807

which the leader and follower repeatedly select their actions in an be a result of deception tactics such as decoy assets for defense
alternating fashion throughout the course of the game. (Salmerón, 2012). For simplicity, we assume that the leader (de-
Sefair and Smith (2016) explore a dynamic shortest path inter- fender) always has accurate information. The interactions between
diction (DSPI) problem, where a defender dynamically decides an the two players are still modeled as a max-min game; however, the
optimal set of arcs to interdict throughout the course of the game. leader aims to maximize the minimum objective of the follower
An interdicted arc increases in cost, as before, and the defender has using actual data, while the follower optimizes their objective us-
a limited number of arcs that (s)he can increase throughout the ing their perceived data. The leader and the follower solve their re-
course of the game. The distinguishing feature of this game is that spective problems using the objective functions of the same form
the defender alternates turns with the attacker (the agent seek- but with different coefficients. This gives rise to a special form of
ing the shortest path). The defender first selects a (possibly empty) bilevel optimization, where the upper-level and lower-level prob-
set of arcs to interdict, and then the attacker chooses a single arc lems only differ by their coefficients.
to traverse. The game iterates in this fashion until the attacker Taking the shortest path interdiction problem (4) as an ex-
reaches its destination node. Unlike most interdiction games, it ample, suppose that the leader’s perception about the cost and
may be uniquely optimal for the defender to interdict fewer arcs interdiction effect on each arc (i, j), cij and dij , respectively, is
than budgeted – in fact, a unique optimal solution may involve in- accurate. The follower, on the other hand, perceives these values
terdicting no arcs at all. Furthermore, the attacker may traverse a differently as c̄i j and d¯i j , due to asymmetric information. Bayrak
cycle in the network, even reusing arcs, at optimality. and Bailey (2008) assume that the follower’s inaccurate perception
The DSPI can be formulated using dynamic programming, is fully known to the leader, and thus formulate the shortest path
where each state consists of the set of arcs interdicted so far and interdiction problem with asymmetric information as a bilevel
the position of the attacker in the network. However, the size optimization problem (see Section 1.3.1). The shortest path net-
of the state space is an exponential function of the defender’s work interdiction problem with asymmetric information can be
interdiction budget. Alternatively, Sefair and Smith (2016) propose modeled as follows:
bounding techniques for the problem by analyzing two variations 
max (ci j + di j xi j )yˆi j , (30)
of the game. The first variation yields a lower bound by supposing x∈X
(i, j )∈A
that interdiction actions “expire” after the attacker’s next move,
meaning that interdicted arcs return to their original costs. Es- where
sentially, this variation serves to reduce the state space of the
 

dynamic program, allowing the problem to be solved in polynomial yˆ ∈ arg min (c̄i j + d¯i j xi j )yi j | (4b), (4c ) .
time. The second variation yields an upper bound by simply reduc- (i, j )∈A
ing the input network to an acyclic network, which again reduces
Information asymmetry also arises in SNIPs where both the
of the state space size to a polynomial function of the network
leader and the follower need to make their decisions before the
size.
realizations of uncertainty involved in the network, i.e., the SNIP
The DSPI is solvable in polynomial time for a fixed interdiction
with a “here-and-now” follower (as opposed to the SNIP with a
budget. This need not be the case for all dynamic interdiction
“wait-and-see” follower, which was mentioned in the previous sec-
games, though, as shown by Sefair and Smith (2017) for a dynamic
tion). In this problem, the risk preferences of both players are rele-
assignment interdiction (DAI) problem. The DAI involves a bipartite
vant to their respective optimization problems, which are likely to
assignment network, in which a set of jobs must be matched to
be heterogenous unless both players are risk neutral (Lei, Shen, &
a set of machines. For the purpose of describing the problem,
Song, 2018). Putting this consideration into the context of stochas-
we say that the follower seeks a minimum-cost assignment, and
tic shortest path interdiction, suppose that the leader and follower
the leader seeks to interdict arcs to maximize the value of a
make their decisions according to their risk measures ρ L ( · ) and
minimum-cost assignment. As before, each arc (representing a
job-to-machine assignment) can be interdicted, resulting in an
ρ F ( · ), respectively, and assume that the follower’s risk preference
is completely known by the leader. Then the leader solves the fol-
increased cost of the corresponding assignment. The game begins
lowing bilevel optimization problem:
with the leader interdicting a set of arcs (as in the DSPI), with the  
follower acting next to assign job 1 to a machine. The leader need 
only consider interdicting assignment arcs corresponding to job min ρL (c˜i j (ξ ) + d˜i j (ξ )xi j )yˆi j , (31)
x∈X
1 in the first iteration. At iteration k ≥ 2 of this game, the leader (i, j )∈A
interdicts some set of arcs corresponding to assigning job k to a
where
machine that has not been assigned to jobs 1, . . . , k − 1, and the    
follower responds by assigning job k to an unassigned machine. 
The total number of arcs that the leader can interdict throughout yˆ ∈ arg min ρF (c˜i j (ξ ) + d˜i j (ξ )xi j )yi j | ( 4b ), ( 4c ) .
the entire game is limited by an upper bound. Like the DSPI, the (i, j )∈A

DAI can again be solved by an exponential state-space dynamic In this setting, information asymmetry gives the defender an
program; however, unlike the DSPI, the DAI remains strongly advantage that can be exploited to better utilize limited defense
NP-hard even if the leader has an interdiction budget of one arc. resources (Salmerón, 2012; Sullivan et al., 2014a). By contrast, we
later consider the case when the attacker (modeled as the fol-
4.4. Network interdiction with asymmetric information lower) holds “private” information that is not completely known
by the defender (for example, the risk preference of the follower
In this section, we discuss cases when assumption A3 is re- ρ F ( · ) may not be completely known to the leader), putting the at-
laxed. This may be simply due to the fact that the two players tacker at an advantage.
have different objective functions in their corresponding optimiza-
tion models. Alternatively, the two players may not have the same 4.5. Network interdiction with incomplete information
perception of their problem data, i.e., assumption A3 is violated
as a result of a violation of assumption A1. Information asymme- In this section, we discuss defender–attacker problems where
try may be inherent in that the defender may have more accu- assumption A1 does not hold, and data pertaining to the network
rate information about the system that they work with, or may interdiction problem is not fully available. Both agents may have to
808 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

learn key parameters of their opponent’s optimization problem ei- interdiction games, the leader refines their incomplete informa-
ther from historical data or from their ongoing interactions. Taking tion about the follower’s problem by learning the cost parameters
the defender’s perspective (as the leader), we assume that the fol- and available resources corresponding to the follower’s problems
lower is fully knowledgeable about their problem, while the leader from the follower’s actions revealed in these games. The leader not
only has partial information about this problem. The leader there- only determines how to optimally allocate interdiction resources
fore makes a robust defensive plan by allowing the follower (the to negatively impact the follower’s objective to the largest ex-
attacker) the advantage of full problem information. tent, but also attempts to make the follower reveal as much infor-
mation as possible so that a more informed interdiction decision
Network interdiction with offline learning. Under assumption A5, can be made in subsequent games. This corresponds to a typical
where only one round of the interdiction game is played, the exploration–exploitation trade-off in online optimization problems
leader makes their interdiction decision based completely on avail- such as the multi-armed bandit problems (Auer, Cesa-Bianchi, &
able historical data. Pay, Merrick, and Song (2019) consider a Fischer, 2002). Borrero et al. (2015) propose a “greedy-and-robust”
stochastic network interdiction problem with a “wait-and-see” fol- policy, where the leader “greedily” optimizes their objective by
lower. The leader is assumed to be a risk-averse decision maker solving a robust optimization problem with an uncertainty set con-
whose preferences can be modeled using expected utility theory structed using the current information available about the follower.
(Aumann, 1962). However, the exact form of the utility function They show that this policy successfully balances exploitation and
is ambiguous to the leader. Pay et al. (2019) assume that histor- exploration in the online optimization setting, eventually matching
ical data on the leader’s past choices is available. The knowledge the optimal interdiction decisions made under perfect information
learned from the data forms an ambiguity set of utility functions after a finite number of rounds.
that are compatible with the leader’s past behaviors. Pay et al. Zheng and Castañón (2012) study dynamic network flow in-
(2019) take a robust optimization perspective by optimizing the in- terdiction games where the attacker has imperfect knowledge of
terdiction decision with respect to the worst-case utility among all the network topology while the network operator (defender) has
utility functions compatible with historical data. perfect information. The attacker can learn about the topology by
An arguably more interesting case is the stochastic network in- monitoring network operations, while the operator observes the
terdiction problem with a “here-and-now” follower, where the fol- attacker’s monitoring actions and chooses to avoid parts of the net-
lower solves an optimization problem under uncertainty according work being monitored by the attacker to hide information from the
to a utility function that characterizes their own risk preference, attacker. After a finite number of stages, the attacker conducts an
but where the follower’s utility function is unknown to the leader. optimal interdiction to maximize the expected network flow dis-
The leader assumes that the follower optimizes their decisions ac- ruption according to their imperfect knowledge about the network
cording to an optimization model (such as the expected utility topology at that point, which is represented by a probability dis-
model), and tries to glean the follower’s private parameters (such tribution over possible states in the information sets. Zheng and
as risk aversion) from the observed follower’s behaviors, which are Castañón (2012) decompose this dynamic game into a sequence of
considered to be optimal solutions to the presumed optimization subgames, each of which is defined starting from an arbitrary ini-
model. The problem of finding the best such estimation is known tial distribution of states in such information sets, and use POMDP
as the inverse optimization problem (Ahuja & Orlin, 2001; Aswani, algorithms to recursively obtain an equilibrium strategy of each
Shen, & Siddiq, 2018; Esfahani, Shafieezadeh-Abadeh, Hanasusanto, subgame.
& Kuhn, 2018). However, from the network interdiction perspec-
tive, assuming that the attacker makes decisions according to an 5. Select applications of network interdiction
optimal solution to a specific model is risky: The follower may
not be solving the optimization model assumed by the leader, or We conclude by discussing newer applications of network
the follower may not have the resources to solve the optimization interdiction and examining how addressing these applications
model exactly. Instead, the leader can construct an ambiguity set will require the development of new interdiction algorithms.
of all possible utility functions that the follower can take in order Section 5.1 examines new challenges of network interdiction ap-
to be compatible with their past behaviors, according to historical plications in cyber-physical system (CPS) security. Section 5.2 dis-
data. The leader then aims at minimizing the highest possible util- cusses network interdiction applications in illicit supply chain
ity value among all utilities contained in this ambiguity set, from network disruption. Section 5.3 summarizes current limitations of
the perspective of protecting against the worst-case scenario. Un- network interdiction models and solution algorithms, and points
fortunately, this problem appears to be computationally challeng- out some future directions of research in this area.
ing, as the inner problem for any fixed interdiction plan becomes a
nonconvex optimization problem. Solution techniques reviewed in 5.1. Cyber-physical system security
Section 3.2 are worth exploring to tackle this problem.
The importance of ensuring the resilience of a large-scale
Network interdiction with online learning. When the interdiction CPS, such as electrical power networks, communication networks,
games are played in multiple rounds (i.e., assumption A5 is re- and transportation networks, cannot be overemphasized. Much
laxed), the leader is provided an opportunity to learn about the of the difficulty associated with these problems stems from the
follower’s optimization problem not only from historical data, but interdependence between different layers of the CPS (including
also from the feedback generated by the follower’s reactions dur- cyber, physical, and cyber-physical) as well as their dynamic
ing these multiple rounds of games. Such a general problem set- interactions. For example, the operation of power grids relies on
ting that involves interplays between optimization and learning real-time control provided by an information network, while infor-
has been studied in the literature under the realm of sequential mation networks also require power supplied from the grid. The
learning and game theory (Cesa-Bianchi & Lugosi, 2006). interdependence of these networks makes them more vulnerable
Borrero, Prokopyev, and Sauré (2015) propose the novel idea of to adversarial attacks than independent networks (Pasqualetti,
formulating sequential network interdiction games in an online op- Dörfler, & Bullo, 2013) because of their larger “attack surfaces.”
timization framework. They assume that the full information ver- Moreover, although classical network interdiction models remain
sion of the problem reduces to the deterministic network interdic- vital in terms of defensive resource allocation at the strategic level,
tion problem. Specifically, by playing multiple rounds of network the dynamic interactions within a CPS impose new challenges for
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 809

applying network interdiction models at the operational level. its devices to balance the expected damage from potential attacks
The complexity of CPS security calls for a holistic framework that with the negative impact brought by false alarms, and the attacker
integrates: (i) game theory – modeling the adversarial relationship acts as the follower, finding the best time to launch the attack that
between the defender and the attacker; (ii) control theory – leads to the highest damage given the detection configuration set
modeling the dynamic interactions between the defender, the by the defender.
attacker and the system; and (iii) system theory – modeling the
interdependence between various layers of the system. Dynamic and interdependence of CPS attacks. Finally, we consider
dynamic CPS defense strategies over time. Recently, cyber and
Defense resource allocation against CPS attacks. We first consider cyber-physical security threats have been growing in both quan-
how existing network interdiction models can be utilized to op- tity and complexity. Defense strategies must be able to adapt to
timize defense resource allocation for CPS security at the strategic real-time information learned from sensors that reveal adversarial
level. Abusorrah, Alabdulwahab, Li, and Shahidehpour (2017) con- emergent behaviors (Strapp & Yang, 2014). Addressing such prob-
sider load redistribution attacks in a smart electric grid (Yuan, Li, & lems requires a holistic framework that incorporates system states
Ren, 2011). Load distribution attacks are a kind of false data injec- and information states, and accordingly optimizes defensive strate-
tion attack that can modify the actual load data di measured by a gies that deal with attackers of heterogeneous types arriving at
sensor in bus i ∈ B to di + di . Assume that the actual system load unknown rates. Most of the existing literature in this area fo-
d is subject to variations that can be described by a polyhedral cuses on static network interdiction models, and does not prop-
uncertainty set D. Under any system load d, let O∗ (d) be the low- erly capture the dynamic nature of interactions between the de-
est possible operational cost without any attack, and let O(w, d) fender and the attackers. Network interdiction models need to be
be the highest possible operational cost induced by a load redis- integrated with models for sequential decision making in an en-
tribution attack given a defense resource allocation w. The value vironment that can itself change dynamically over time, such as
O∗ (d) can be computed by solving an optimal power flow problem MDP models (Gutin, Kuhn, & Wiesemann, 2014). The situation is
min p ∈ P(d) F(p), which minimizes total operational cost F(p) over all further complicated when neither player has complete informa-
feasible power flows p that satisfy certain physical constraints P(d) tion about the other player’s payoff function, feasible action space,
based on the load distribution d. Computing O(w, d) involves solv- etc., as discussed in Section 4.5, in which case models such as
ing a bilevel problem: POMDP (Baykal-Gürsoy et al., 2014; Zheng & Castañón, 2012) and
  repeated Bayesian games (Liu, Comaniciu, & Man, 2006) should be
max min F (p ) , (32) integrated. See Etesami and Başar (2019) for a recent survey on
d∈C (w,d ) p∈P˜(d,d ) dynamic security games arising in CPS.
where C (w, d ) defines the possible load redistribution attacks that
5.2. Interdicting illicit supply networks
can be done given defense allocation w and system load d, and
P˜(d, d ) is the set of feasible power flow under the corrupted
In this section, we discuss new applications of network inter-
data. Model (32) can be reformulated as a single-level maxi-
diction problems to disrupting illicit supply chain networks. Tra-
mization problem by using the standard “dualize-and-combine”
ditionally, these applications focus on smuggling illegal materials
approach.
such as nuclear weapons (McLay, Lloyd, & Niman, 2011; Morton
Abusorrah et al. (2017) then model the defender’s decision-
et al., 2007) through a given network. The smugglers (attackers)
making from a robust optimization perspective, and consider de-
maximize the amount of illegal materials sent from an origin to a
fending against the worst case scenario in terms of the difference
destination. This can be formulated as a maximum flow problem.
between O(w, d) and O∗ (d) among all possible system load varia-
The defender, on the other hand, optimizes their interdiction deci-
tions d ∈ D:
sions according to a budget constraint to diminish the optimal flow
min max {O(w, d ) − O∗ (d )}, (33) value.
w∈S d∈D
New applications of network interdiction problems on illicit
where S gives the defense resource allocation constraints, includ- supply network disruption focus on modeling the interdepen-
ing, e.g., financial budget constraints and computing resource con- dence between the interdiction decisions defined on the network.
straints, limiting the number of security checkpoints that can be Malaviya, Rainwater, and Sharkey (2019) consider a problem mo-
placed in the CPS network. Algorithms described in Section 3.2 can tivated by city-level illegal drug enforcement applications. They
be applied to solve the min-max MIP problem (33). model an information network connecting drug dealers at each
hierarchical level. This information network is interconnected with
Attack detection in CPS. Attack detection is usually ignored by clas- a physical network, communicating information on drug dealers’
sical network interdiction models, which assume the existence of actions and providing feedback to them. Law enforcement officials
a single attacker whose actions can be fully observed and eval- make interdiction decisions over multiple periods that follow a
uated. In modern CPS, various types of sensors are distributed certain logical sequence: local drug dealers need to be captured
across the network to collect data, and anomalous attacks are de- (interdicted) first in order to provide evidence/information for
tected and identified by statistical or machine-learning algorithms interdicting a higher-level drug dealer at a later time. Baycik,
based on collected sensor data. Due to the variety of attacks that Sharkey, and Rainwater (2018) model the interdependent infor-
could be launched on a CPS, attack detection devices and algo- mation network and physical network as a multi-layered network,
rithms are imperfect. These devices can be configured with re- and consider interdictions made on both networks simultaneously,
spect to their detection sensitivity, allowing the CPS operator to with applications in both city-level illegal drug disruption and CPS
find the appropriate balance between the expected time to de- security.
tection and false alarm rate. Ghafouri, Abbas, Laszka, Vorobeychik, Another prominent challenge for network interdiction with
and Koutsoukos (2016) discuss the problem of optimizing detection integrated information and physical networks is that key struc-
thresholds that minimize expected damage incurred by detection tures in the information networks used by the attackers are
delay, while maintaining the false alarm rate to be within a toler- intentionally hidden from the defender. Konrad, Trapp, Palmbach,
able limit. The attack detection problem itself can be seen as an and Blom (2017) examine such problems in human trafficking
interdiction problem: the defender acts as the leader, configuring networks, which include hidden victims and covert traffickers.
810 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811

The defender may have to commit defense resources without full Bertsimas, D., & Thiele, A. (2006). Robust and data-driven optimization: Modern de-
knowledge of the information network. It is therefore critical to cision-making under uncertainty. In M. P. Johnson, B. Norman, & N. Secomandi
(Eds.), Tutorials in operations research: Models, methods, and applications for in-
gather incomplete information piece by piece from various sources novative decision making (pp. 95–122). Catonsville, MD: INFORMS.
such as social media, and make robust interdiction decisions by Bodur, M., Dash, S., Günlük, O., & Luedtke, J. (2016). Strengthened Benders cuts for
effectively hedging against uncertainty. These challenges further stochastic integer programs with continuous recourse. INFORMS Journal on Com-
puting, 29(1), 77–91.
motivate emerging interdiction studies described in Section 4, Borrero, J. S., Prokopyev, O. A., & Sauré, D. (2015). Sequential shortest path interdic-
such as network interdiction with incomplete information in a tion with incomplete information. Decision Analysis, 13(1), 68–98.
dynamic environment. Brown, G. G., Carlyle, W. M., Harney, R. C., Skroch, E. M., & Wood, R. K. (2009).
Interdicting a nuclear-weapons project. Operations Research, 57(4),
866–877.
5.3. Current limitations and some future directions of research Brown, G. G., Carlyle, W. M., Salmerón, J., & Wood, R. K. (2006). Defending critical
infrastructure. Interfaces, 36(6), 530–544.
Caprara, A., Carvalho, M., Lodi, A., & Woeginger, G. J. (2016). Bilevel knap-
In the following, we identify some current limitations of mod- sack with interdiction constraints. INFORMS Journal on Computing, 28(2), 319–
eling and solution methodologies for addressing challenges arising 333.
Cesa-Bianchi, N., & Lugosi, G. (2006). Prediction, learning, and games. Cambridge, UK:
from new network interdiction problem variants.
Cambridge University Press.
First, the computational efficiency of optimizing adaptive de- Church, R. L., Scaparra, M. P., & Middleton, R. S. (2004). Identifying critical infras-
fense strategies is a bottleneck for successfully applying them to tructure: The median and covering facility interdiction problems. Annals of the
Association of American Geographers, 94(3), 491–502.
large-scale systems such as those pertaining to electric power
Codato, G., & Fischetti, M. (2006). Combinatorial Benders’ cuts for mixed-integer
grids. It is crucial to develop next-generation computationally ef- linear programming. Operations Research, 54(4), 756–766.
ficient algorithms and robust computational infrastructure that en- Cormican, K. J., Morton, D. P., & Wood, R. K. (1998). Stochastic network interdiction.
able these algorithms to compute near-optimal strategies. Operations Research, 46(2), 184–197.
Dantzig, G. B., & Fulkerson, D. R. (1955). On the max flow min cut theorem of net-
Second, the complexity of a holistic modeling framework for works. Technical Report, P-826. Santa Monica, CA: The RAND Corporation.
defending large-scale interdependent systems calls for a multi- Delage, E., & Iancu, D. A. (2015). Robust multistage decision making. In D. M. Ale-
agent framework with varying levels of autonomy among multiple man, & A. C. Thiele (Eds.), Tutorials in operations research: The operations research
revolution (pp. 20–46). Catonsville, MD: INFORMS.
defenders. The multi-agent framework should characterize how de- Dempe, S., Kalashnikov, V. V., Pérez-Valdés, G. A., & Kalashnykova, N. (2015). Bilevel
fenders learn and adapt from their local environment with limited programming problems. Heidelberg: Springer.
data, and share information with the centralized decision maker DeNegre, S. T., & Ralphs, T. K. (2009). A branch-and-cut algorithm for integer bilevel
linear programs. In J. W. Chinneck, B. Kristjansson, & M. J. Saltzman (Eds.), Op-
(if any exists) and other local defenders via limited communica- erations research and cyber-infrastructure (pp. 65–78). New York: Springer.
tion (Leitao et al., 2016; Sreekumaran, Hota, Liu, Uhan, & Sun- Elias, P., Feinstein, A., & Shannon, C. E. (1956). A note on the maximum flow through
daram, 2015). a network. IRE Transactions on Information Theory, 2(4), 117–119.
Esfahani, P. M., Shafieezadeh-Abadeh, S., Hanasusanto, G. A., & Kuhn, D. (2018).
Third, most existing defense strategies are reactive in nature.
Data-driven inverse optimization with imperfect information. Mathematical Pro-
As cyber and cyber-physical security threats are growing fast in gramming, 167(1), 191–234.
both quantity and variety, existing reactive defense strategies may Etesami, S. R., & Başar, T. (2019). Dynamic games in cyber-physical security: An
overview. Dynamic Games and Applications. doi:10.1007/s13235- 018- 00291- y.
soon reach their limits in mitigating these threats. New defense
Fischetti, M., Ljubic, I., Monaci, M., & Sinnl, M. (2019). Interdiction games and mono-
strategies, such as proactively disrupting the social–information tonicity, with application to knapsack problems. INFORMS Journal on Computing,
networks that connect adversarial organizations and agents should 31(2), 390–410.
be developed. Ford, L. R., & Fulkerson, D. R. (1956). Maximal flow through a network. Canadian
Journal of Mathematics, 8, 399–404.
Fulkerson, D. R., & Harding, G. C. (1977). Maximizing minimum source-sink
References path subject to a budget constraint. Mathematical Programming, 13(1), 116–
118.
Abusorrah, A., Alabdulwahab, A., Li, Z., & Shahidehpour, M. (2017). Minimax-regret Ghafouri, A., Abbas, W., Laszka, A., Vorobeychik, Y., & Koutsoukos, X. (2016). Op-
robust defensive strategy against false data injection attacks. IEEE Transactions timal thresholds for anomaly-based intrusion detection in dynamical environ-
on Smart Grid, 10(2), 2068–2079. ments. In Q. Zhu, T. Alpcan, E. Panaousis, M. Tambe, & W. Casey (Eds.), Pro-
Ahuja, R. K., & Orlin, J. B. (2001). Inverse optimization. Operations Research, 49(5), ceedings of the international conference on decision and game theory for security
771–783. (pp. 415–434). New York: Springer.
Alarie, S., Audet, C., Jaumard, B., & Savard, G. (2001). Concavity cuts for disjoint bi- Ghare, P. M., Montgomery, D. C., & Turner, W. C. (1971). Optimal interdiction policy
linear programming. Mathematical Programming, 90(2), 373–398. for a flow network. Naval Research Logistics Quarterly, 18(1), 37–45.
Altner, D. S., Ergun, O., & Uhan, N. A. (2010). The maximum flow network inter- Goldberg, N. (2017). Non-zero-sum nonlinear network path interdiction with an
diction problem: Valid inequalities, integrality gaps, and approximability. Oper- application to inspection in terror networks. Naval Research Logistics, 64(2),
ations Research Letters, 38(1), 33–38. 139–153.
Assimakopoulos, N. (1987). A network interdiction model for hospital infection con- Golden, B. (1978). A problem in network interdiction. Naval Research Logistics Quar-
trol. Computers in Biology and Medicine, 17(6), 413–422. terly, 25(4), 711–713.
Aswani, A., Shen, Z.-J., & Siddiq, A. (2018). Inverse optimization with noisy data. Guan, P., He, M., Zhuang, J., & Hora, S. (2017). Modeling a multi-target attacker–de-
Operations Research, 66(3), 870–892. fender game with budget constraints. Decision Analysis, 14(2), 87–107.
Atamtürk, A., Deck, C., & Jeon, H. (2019). Successive quadratic upper-bounding for Günlük, O., & Pochet, Y. (2001). Mixing mixed-integer inequalities. Mathematical
discrete mean-risk minimization and network interdiction. BCOL Research Report Programming, 90(3), 429–457.
17.05, UC Berkeley (in press). Gutin, E., Kuhn, D., & Wiesemann, W. (2014). Interdiction games on Markovian PERT
Atamtürk, A., Nemhauser, G. L., & Savelsbergh, M. W. (20 0 0). The mixed vertex pack- networks. Management Science, 61(5), 999–1017.
ing problem. Mathematical Programming, 89(1), 35–53. Harris, T. E., & Ross, F. S. (1955). Fundamentals of a method for evaluating rail net
Auer, P., Cesa-Bianchi, N., & Fischer, P. (2002). Finite-time analysis of the multiarmed capacities. Research Memorandum, RM-1573. Santa Monica, CA: The RAND Cor-
bandit problem. Machine Learning, 47(2–3), 235–256. poration.
Aumann, R. J. (1962). Utility theory without the completeness axiom. Econometrica: Israeli, E., & Wood, R. K. (2002). Shortest-path network interdiction. Networks, 40(2),
Journal of the Econometric Society, 30(3), 445–462. 97–111.
Ball, M. O., Golden, B. L., & Vohra, R. V. (1989). Finding the most vital arcs in a Janjarassuk, U., & Linderoth, J. (2008). Reformulation and sampling to solve a
network. Operations Research Letters, 8(2), 73–76. stochastic network interdiction problem. Networks, 52(3), 120–132.
Baycik, N. O., Sharkey, T. C., & Rainwater, C. E. (2018). Interdicting layered physical Kall, P., & Wallace, S. W. (1994). Stochastic programming. Chichester, UK: John Wiley
and information flow networks. IISE Transactions, 50(4), 316–331. and Sons.
Baykal-Gürsoy, M., Duan, Z., Poor, H. V., & Garnaev, A. (2014). Infrastructure security Konrad, R. A., Trapp, A. C., Palmbach, T. M., & Blom, J. S. (2017). Overcoming hu-
games. European Journal of Operational Research, 239(2), 469–478. man trafficking via operations research and analytics: Opportunities for meth-
Bayrak, H., & Bailey, M. D. (2008). Shortest path network interdiction with asym- ods, models, and applications. European Journal of Operational Research, 259(2),
metric information. Networks, 52(3), 133–140. 733–745.
Ben-Tal, A., El Ghaoui, L., & Nemirovski, A. (2009). Robust optimization. Princeton Lawler, E. L. (1976). Combinatorial optimization, networks and matroids. Dover.
series in applied mathematics. Princeton, NJ: Princeton University Press. Lei, X., Shen, S., & Song, Y. (2018). Stochastic maximum flow interdiction prob-
Bertsimas, D., & Sim, M. (2004). The price of robustness. Operations Research, 52(1), lems under heterogeneous risk preferences. Computers & Operations Research,
35–53. 90, 97–109.
 J.C. Smith and Y. Song / European Journal of Operational Research 283 (2020) 797–811 811

Leitao, P., Karnouskos, S., Ribeiro, L., Lee, J., Strasser, T., & Colombo, A. W. (2016). Shapiro, A., Dentcheva, D., & Ruszczynski, A. (2009). Lectures in stochastic program-
Smart agents in industrial cyber-physical systems. Proceedings of the IEEE, ming: modeling and theory. Philadelphia, PA: SIAM.
104(5), 1086–1101. Sherali, H. D., & Shetty, C. M. (1980). A finitely convergent algorithm for bilinear
Lemaréchal, C., Nemirovski, A., & Nesterov, Y. (1995). New variants of bundle meth- programming problems using polar cuts and disjunctive face cuts. Mathematical
ods. Mathematical Programming, 69(1–3), 111–147. Programming, 19, 14–31.
Lim, C., & Smith, J. C. (2007). Algorithms for discrete and continuous multicommod- Smith, J. C., & Ahmed, S. (2011). Introduction to robust optimization. In J. J. Cochran
ity flow network interdiction problems. IIE Transactions, 39(1), 15–26. (Ed.), Wiley encyclopedia of operations research and management science. Hobo-
Liu, Y., Comaniciu, C., & Man, H. (2006). A Bayesian game approach for intrusion ken, NJ: Wiley.
detection in wireless ad hoc networks. Proceeding from the 2006 workshop on Smith, J. C., Lim, C., & Alptekinoglu, A. (2009). New product introduction against a
game theory for communications and networks. Pisa, Italy: ACM. predator: A Bilevel Mixed-integer Programming approach. Naval Research Logis-
Lozano, L., & Smith, J. C. (2017). A value-function-based exact approach for tics, 56(8), 714–729.
the bilevel mixed-integer programming problem. Operations Research, 65(3), Smith, J. C., Lim, C., & Sudargho, F. (2007). Survivable network design under op-
768–786. timal and heuristic interdiction scenarios. Journal of Global Optimization, 38(2),
Magnanti, T. L., & Wong, R. T. (1981). Accelerating Benders decomposition: Algo- 181–199.
rithmic enhancement and model selection criteria. Operations Research, 29(3), Song, Y., & Shen, S. (2016). Risk-averse shortest path interdiction. INFORMS Journal
464–484. on Computing, 28(3), 527–539.
Mahdavi Pajouh, F., Boginski, V., & Pasiliao, E. L. (2014). Minimum vertex blocker Sreekumaran, H., Hota, A. R., Liu, A. L., Uhan, N. A., & Sundaram, S. (2015). Multi-
clique problem. Networks, 64(1), 48–64. agent decentralized network interdiction games. arXiv preprint:1503.01100.
Malaviya, A., Rainwater, C., & Sharkey, T. C. (2012). Multi-period network interdic- von Stackelberg, H. (1952). The theory of the market economy. London: William
tion problems with applications to city-level drug enforcement. IIE Transactions, Hodge and Co.
44(5), 368–380. Strapp, S., & Yang, S. J. (2014). Segmenting large-scale cyber attacks for online be-
McLay, L. A., Lloyd, J. D., & Niman, E. (2011). Interdicting nuclear material on havior model generation. In W. G. Kennedy, N. Agarwal, & S. J. Yang (Eds.), Pro-
cargo containers using knapsack problem models. Annals of Operations Research, ceedings of the international conference on social computing, behavioral-cultural
187(1), 185–205. modeling, and prediction (pp. 169–177). New York: Springer.
McMasters, A., & Mustin, T. M. (1970). Optimal interdiction of a supply network. Sullivan, K. M., Morton, D. P., Pan, F., & Smith, J. C. (2014a). Securing a border under
Naval Research Logistics Quarterly, 17, 261–268. asymmetric information. Naval Research Logistics, 61(2), 91–100.
Mitsos, A. (2010). Global solution of nonlinear mixed-integer bilevel programs. Jour- Sullivan, K. M., Smith, J. C., & Morton, D. P. (2014b). Convex hull representation of
nal of Global Optimization, 47(4), 557–582. the deterministic bipartite network interdiction problem. Mathematical Program-
Morton, D. P., Pan, F., & Saeger, K. J. (2007). Models for nuclear smuggling interdic- ming, 145(1–2), 349–376.
tion. IIE Transactions, 39(1), 3–14. Tang, Y., Richard, J.-P. P., & Smith, J. C. (2016). A class of algorithms for mixed-in-
Morton, D. P., & Wood, R. K. (1999). Restricted-recourse bounds for stochastic linear teger bilevel min–max optimization. Journal of Global Optimization, 66(2),
programming. Operations Research, 47(6), 943–956. 225–262.
Nemhauser, G. L., & Wolsey, L. A. (1988). Integer and combinatorial optimization. New Towle, E., & Luedtke, J. (2018). New solution approaches for the maximum-reliabil-
York: Wiley. ity stochastic network interdiction problem. Computational Management Science,
Nguyen, K. T. (2016). Reverse 1-center problem on weighted trees. Optimization, 15(3–4), 455–477.
65(1), 253–264. Washburn, A., & Wood, R. K. (1995). Two-person zero-sum games for network in-
Pan, F., & Morton, D. P. (2008). Minimizing a stochastic maximum-reliability path. terdiction. Operations Research, 43(2), 243–251.
Networks, 52, 111–119. Wollmer, R. D. (1964). Removing arcs from a network. Operations Research, 12(6),
Pasqualetti, F., Dörfler, F., & Bullo, F. (2013). Attack detection and identification in cy- 934–940.
ber-physical systems. IEEE Transactions on Automatic Control, 58(11), 2715–2729. Wood, R. K. (1993). Deterministic network interdiction. Mathematical and Computer
Pay, B. S., Merrick, J. R. W., & Song, Y. (2019). Stochastic network interdiction with Modelling, 17(2), 1–18.
incomplete preference. Networks, 73(1), 3–22. Wu, X., & Conejo, A. J. (2017). An efficient tri-level optimization model for electric
Phillips, C. A. (1993). The network inhibition problem. In Proceedings of the twen- grid defense planning. IEEE Transactions on Power Systems, 32(4), 2984–2994.
ty-fifth annual ACM symposium on the theory of computing (pp. 776–785). New Xu, P., & Wang, L. (2014). An exact algorithm for the bilevel mixed integer linear
York: ACM. programming problem under three simplifying assumptions. Computers and Op-
Powell, R. (2007). Defending against terrorist attacks with limited resources. Ameri- erations Research, 41(1), 309–318.
can Political Science Review, 101(3), 527–541. Yuan, W., Wang, J., Qiu, F., Chen, C., Kang, C., & Zeng, B. (2016). Robust optimiza-
Ratliff, H. D., Sicilia, G. T., & Lubore, S. H. (1975). Finding the n most vital links in tion-based resilient distribution network planning against natural disasters. IEEE
flow networks. Management Science, 21(5), 531–539. Transactions on Smart Grid, 7(6), 2817–2826.
Rocco, C. M., & Ramirez-Marquez, J. E. (2010). A bi-objective approach for short- Yuan, W., Zhao, L., & Zeng, B. (2014). Optimal power grid protection through a de-
est-path network interdiction. Computers and Industrial Engineering, 59(2), 232– fender–attacker–defender model. Reliability Engineering and System Safety, 121,
240. 83–89.
Rocco, C. M., Ramirez-Marquez, J. E., & Salazar, D. E. (2010). Bi and tri-objective Yuan, Y., Li, Z., & Ren, K. (2011). Modeling load redistribution attacks in power sys-
optimization in the deterministic network interdiction problem. Reliability Engi- tems. IEEE Transactions on Smart Grid, 2(2), 382–390.
neering and System Safety, 95(8), 887–896. Zeng, B., & Zhao, L. (2013). Solving two-stage robust optimization problems using
Royset, J. O., & Wood, R. K. (2007). Solving the bi-objective maximum-flow net- a column-and-constraint generation method. Operations Research Letters, 41(5),
work-interdiction problem. INFORMS Journal on Computing, 19(2), 175–184. 457–461.
Saharidis, G. K., & Ierapetritou, M. G. (2009). Resolution method for mixed integer Zenklusen, R. (2010). Matching interdiction. Discrete Applied Mathematics, 158(15),
bi-level linear problems based on decomposition technique. Journal of Global 1676–1690.
Optimization, 44(1), 29–51. Zhao, L., & Zeng, B. (2013). Vulnerability analysis of power grids with line switching.
Salmerón, J. (2012). Deception tactics for network interdiction: A multiobjective ap- IEEE Transactions on Power Systems, 28(3), 2727–2736.
proach. Networks, 60(1), 45–58. Zheng, J., & Castañón, D. A. (2012). Dynamic network interdiction games with im-
Salmerón, J., Wood, K., & Baldick, R. (2009). Worst-case interdiction analysis of perfect information and deception. In Proceedings of the fifty-first IEEE conference
large-scale electric power grids. IEEE Transactions on Power Systems, 24(1), on decision and control (CDC) (pp. 7758–7763). Maui, HI: IEEE.
96–104. Zhuang, J., & Bier, V. M. (2007). Balancing terrorism and natural disasters—Defensive
Sefair, J. A., & Smith, J. C. (2016). Dynamic shortest-path interdiction. Networks, strategy with endogenous attacker effort. Operations Research, 55(5), 967–991.
68(4), 315–330.
Sefair, J. A., & Smith, J. C. (2017). Exact algorithms and bounds for the dynamic
assignment interdiction problem. Naval Research Logistics, 64(5), 373–387.