Deceitful Attacks in Security Games

Thanh Hong Nguyen, Michael P. Wellman, Arunesh Sinha

Abstract: Given recent applications of defender-attacker Stackelberg Security Games in real-world domains such as wildlife protection, a majority of research has focused on addressing uncertainties regarding the attacker in these games based on the exploitation of attack data. However, there is an important challenge of deceitful attacks; the attacker can manipulate his attacks to mislead the defender, leading her to conduct ineffective patrolling strategies. In this work, we focus on addressing this challenge while providing the following main contributions. First, we introduce a new game model with uncertainty about the attacker type and repeated interactions between the players. In our game model, the defender attempts to collect attack data over time to learn about the attacker type while the attacker aims at playing deceitfully. Second, based on the new game model, we propose new game-theoretic algorithms to compute optimal strategies for both players. Third, we present preliminary experiment results to evaluate our proposed algorithms, showing that our defense solutions can effectively address deceitful attacks.