Empirical Comparisons of Descriptive Multi-objective Adversary Models in Stackelberg Security Games

Abstract

Stackelberg Security Games (SSG) have been used to model defender- attacker relationships for analyzing real-world security resource allocation problems. Research has focused on generating algorithms that are optimal and efficient for defenders, based on a presumed model of adversary choices. However, relatively less has been done descriptively to investigate how well those models capture adversary choices and psychological assumptions about adversary decision making. Using data from three experiments, including over 1000 human subjects playing over 25000 games, this study evaluates adversary choices by comparing 9 adversary models both nomothetically and ideographically in a SSG setting. We found that participants tended to be consistent with utility maximization and avoid a target with high probability of being protected even if the reward or expected value of that target is high. It was also found in two experiments that adversary choices were dependent on the defender’s payoffs, even after accounting for attacker’s own payoffs.