Arunesh Sinha

"What's in a Game? An intelligent and adaptive approach to security"

(Vortrag im Rahmen der "MPI Distinguished Lecture Series" in Kooperation mit dem Fachbereich Informatik)

Understanding the complex defender-adversary interaction in any adversarial interaction allows for the design of intelligent and adaptive defense. Game theory is a natural model for such multi-agent interaction. However, significant challenges need to be overcome in order to apply game theory in practice. In this talk, I will present my work on addressing two such challenges: scalability and learning adversary behavior. First, I will present a game model of screening of passengers at airports and a novel optimization approach based on randomized allocation and disjunctive programming techniques to solve large instances of the problem. Next, I will present an approach that learns adversary behavior and then plans optimal defensive actions, thereby bypassing standard game-theoretic assumptions such as rationality. However, a formal Probably Approximately Correct (PAC) model analysis of the learning module in such an approach reveals possible conditions under which learning followed by optimization can produce sub-optimal results. This emphasizes the need of formal compositional reasoning when using learning in large systems.

The airport screening work was done in collaboration with the Transport Security Administration in USA. The approach of learning adversary behavior was applied for predictive policing in collaboration with University of Southern California (USC) police, and is being tested on the USC campus.