TL;DR
This paper introduces a novel hybrid model for homeland security resource allocation that accounts for attackers who might be either fully strategic or non-strategic. It demonstrates that game-theoretic approaches, which model strategic interactions, are generally more robust and lead to lower expected losses for the defender, even when the attacker's behavior is uncertain.
What Problem Does This Paper Address?
Homeland security agencies face the immense challenge of allocating limited resources to protect numerous potential targets from terrorist attacks. Traditional risk analysis often treats threats as static or random. While game theory offers a powerful framework for modeling strategic adversaries, it typically assumes attackers are fully rational and possess significant information. This assumption may not always hold true, as terrorists might act non-strategically due to various reasons. Conversely, ignoring the strategic nature of terrorists can lead to suboptimal defense allocations. This paper bridges this gap by developing a model that explicitly considers the possibility of a "partially strategic" attacker, allowing for a spectrum of attacker behaviors from fully strategic to fully non-strategic. This allows for a more nuanced understanding of how to best allocate defensive resources in the face of uncertainty about the adversary's decision-making process.
Methods
The core of this paper is a sequential defender-attacker game model. The defender (government) first allocates a fixed budget of defensive resources across multiple targets. Subsequently, the attacker (terrorist) observes this allocation and chooses which targets to attack. The novelty lies in modeling the attacker as having a probability q of being strategic and 1-q of being non-strategic.
- Strategic Attacker: A strategic attacker's decision is endogenous. They will choose to attack targets that maximize their expected payoff (which, in this model, corresponds to maximizing the defender's expected loss). Their attack probabilities are determined by the defender's resource allocation.
- Non-Strategic Attacker: A non-strategic attacker's behavior is exogenous. Their attack probabilities are pre-defined and do not adapt to the defender's resource allocation.
The model formulates the defender's objective as minimizing total expected loss, considering both types of attackers. It then derives conditions for a subgame perfect Nash equilibrium and develops an algorithm to find optimal defense resource allocations. A key aspect is the analysis of "robustness," which measures how well the game-theoretic approach performs under different assumptions about the attacker's strategy and the defender's beliefs.
Datasets
The paper utilizes data from the U.S. Department of Homeland Security (DHS) and other sources to represent:
- Target Valuations (xᵢ): Expected property losses for 47 U.S. urban areas, sourced from Willis et al. (2005).
- Defense Budget (C): The total available defense budget is set to the FY 2004 UASI Grant Allocations ($675 million).
- Cost-Effectiveness of Defense (λᵢ): Represented by a parameter k in the success probability function, assumed to be homogeneous across targets.
Key Insights and Ideas
- Game Theory is Robust: The most striking insight is that game-theoretic models, which assume strategic attackers, are generally preferred even when the probability of a non-strategic attacker is high (e.g., > 50%). This suggests that assuming a more capable adversary is often a safer bet.
- "Conservative" Defense: Assuming the attacker is strategic when they might be non-strategic leads to a "conservative" defense strategy that incurs less expected loss compared to assuming the attacker is non-strategic when they might be strategic.
- Defender's Beliefs Matter, But Less Than Reality: While the defender's beliefs about the attacker's strategy influence their optimal allocation, the actual behavior of the attacker (strategic vs. non-strategic) has a more profound impact on the resulting expected loss.
- Targeting High-Value Assets: The model generally supports allocating resources to high-value targets, especially when facing a strategic attacker. However, the specific allocation can shift based on the probability of a non-strategic attacker.
Key Results
- Equilibrium Existence: The paper proves that a unique subgame perfect Nash equilibrium exists for the formulated game.
- Preference Threshold (T): A threshold T is identified such that if the probability of a non-strategic attacker (1-q) is less than or equal to T, game-theoretic models are preferred. This threshold is often found to be significantly greater than 0.5, indicating a broad preference for game-theoretic approaches.
- Budget Allocation Sensitivity: Optimal defense budget allocations are shown to be relatively insensitive to the probability of a non-strategic attacker, especially when this probability is small, the cost-effectiveness of defense is high, or the number of top-tier targets is large.
- Robustness Measure (d): The difference in expected loss between game-theoretic and non-game-theoretic models (robustness measure d) is generally positive, indicating the superiority of game-theoretic approaches in most scenarios, particularly when the budget is low or high, defense cost-effectiveness is high, or the attacker is more likely to be strategic.
Limitations and Future Work
- Assumptions: The model assumes perfect information about the attacker's behavioral types (strategic vs. non-strategic) and their probabilities. It also assumes the defender and attacker share the same target valuations. The success probability function is assumed to be exponential.
- Computational Complexity: While an algorithm is provided, scaling to very large numbers of targets might still be a challenge.
- Future Directions: The authors suggest developing more efficient algorithms for larger-scale problems, incorporating more general success probability functions, exploring multi-period attacks, and considering networked targets that can be impacted by a single attack. They also propose incorporating more sophisticated attacker utility functions that account for uncertainty and multiple attributes.
Summary updated