Developing the Science and Applications of Security Games: Machine Learning, Uncertainty and Preference Elicitation in Game Theory for Security

Principal Investigator: 
Performance Period: 
July 2015 to June 2016
Project Status: 
In Progress
Commercialization Status: 
Project Keywords: 
game theory
security games
machine learning
mathematical programming
In the past, we successfully founded the research area of security games. This research focuses on using game theory – often defender-attacker Stackelberg games -- to optimize the use of limited security resources, accounting for adversaries who conduct surveillance and then react; it also led to one of the most successful set of applications of computational game theory. Among our past application successes include ARMOR, to randomize vehicle checkpoints and canine unit patrols at the Los Angeles International Airport (LAX); IRIS, deployed by the Federal Air Marshals Service (FAMS); PROTECT, deployed by the US Coast Guard in Boston, Houston, Los Angeles, and New York for port protection; and several others. New applications have recently emerged however, including US Coast Guard’s protection of fisheries, protection against crimes in the LA Metro System, US Coast Guard’s interdiction of drugs and protection of wildlife against poaching (which isn’t strictly a DHS mission but is in the same family of security games, and note that illegal poaching funds terrorist groups like Al-Shabab). These new set of applications bring to fore some fundamental new research challenges arising from key novel characteristics such as: (i) these are games involving repeated interaction with adversaries; (ii) there is not a single but an entire heterogeneous population of adversaries; (iii) the adversaries may act more opportunistically rather than fully strategically; (iv) there is significant adversary data available. The original security game model doesn’t apply as a result, since it assumes a single shot game with a highly strategic adversary, with limited or no data about adversaries.   We propose several directions to address these new challenges. The repeated nature of the interactions and the presence of significant data, allows for machine learning techniques; the opportunistic nature of crime requires formulating opportunistic security games; the presence of data (if not in significant amounts) about adversaries leads to solving Robust Bayesian games; finally we address preference elicitation techniques, a universal challenge applicable in many domains.