To Handle, to Learn and to Manipulate the Attacker's (Uncertain) Payoffs in Security Games: Doctoral Consortium

Stackelberg security games (SSGs) are now established as a powerful tool in security domains. In order to compute the optimal strategy for the defender in SSG model, the defender needs to know the attacker's preferences over targets so that she can predict how the attacker would react under a certain defender strategy. Uncertainty over attacker preferences may cause the defender to suffer large losses. My thesis focuses on uncertainty in attacker preferences: such uncertainty may arise because of uncertainty over attacker's risk attitude or uncertainty over true attacker payoffs. To that end, the first part of my thesis focuses on risk-averse attackers. Extensive studies show that the attackers in some domains are in fact risk-averse rather than risk-neutral, which has never been taken into account in previous security game literatures. To handle the attacker's risk aversion attitude in viewing payoffs, I develop an algorithm to compute the defender's robust strategy against an uncertain risk-averse attacker since the defenders are also uncertain about the degree of attacker's risk aversion. The second part of my thesis focuses on learning attacker payoffs. More specifically, the concept of SSGs has also been applied to the domain of protecting natural resources, where the attacker "attacks"(illegally extracts natural resources) frequently, which reveals his preference over different targets. Based on this concept, I develop the algorithm for the defender to learn target values from the attacker's actions and then use this information to better plan her strategy.