%0 Thesis %D 2017 %T When AI helps Wildlife Conservation: Learning Adversary Behaviors in Green Security Games %A Kar, Debarun %X Whereas previous real-world game-theoretic applications in security focused on protection of critical infrastructure in the absence of past attack data, more recent work has focused on datadriven security and sustainability applications for protecting the environment, including forests, fish and wildlife. One key challenge in such “Green Security Game” (GSG) domains is to model the adversary’s decision making process based on available attack data. This thesis, for the first time, explores the suitability of different adversary behavior modeling approaches in such domains that differ in the type and amount of historical data available. The first contribution is to provide a detailed comparative study, based on actual human subject experiments, of competing adversary behavior models in domains where attack data is available in plenty (e.g., via a large number of sensors). This thesis demonstrates a new human behavior model, SHARP, which mitigates the limitations of previous models in three key ways. First, SHARP reasons based on successes or failures of the adversary’s past actions to model adversary adaptivity. Second, SHARP reasons about similarity between exposed and unexposed areas of the attack surface to handle the adversary’s lack of exposure to enough of the attack surface. Finally, SHARP integrates a non-linear probability weighting function to capture the adversary’s true weighting of probabilities.The second contribution relates to domains requiring predictions over a large set of targets by learning from limited (and in some cases, noisy) data. One example dataset on which we demonstrate our approaches to handle such challenges is a real-world poaching dataset collected over a large geographical area at the Queen Elizabeth National Park in Uganda. This data is too sparse to construct a detailed model. The second contribution of this thesis delivers a surprising result by presenting an adversary behavior modeling system, INTERCEPT, which is based on an ensemble of decision trees (i) that effectively learns and predicts poacher attacks based on limited noisy attack data over a large set of targets, and (ii) has fast execution speed. This has led to a successful month-long test of INTERCEPT in the field, a first for adversary behavior modeling applications in the wildlife conservation domain. Finally, for the my third contribution, we examine one common assumption in adversary behavior modeling that the adversary perfectly observes the defender’s randomized protection strategy. However, in domains such as wildlife conservation, the adversary only observes a limited sequence of defender patrols and forms beliefs about the defender’s strategy. In the absence of a comparative analysis and a principled study of the strengths and weaknesses of belief models, no informed decision could be made to incorporate belief models in adversary behavior models such as SHARP and INTERCEPT. This thesis provides the first-of-its-kind systematic comparison of existing and new proposed belief models and demonstrates based on human subjects experiments data that identifying heterogeneous belief update behavior is essential in making effective predictions. We also propose and evaluate customized models for settings that differ in the type of belief data available and quantify the value of having such historical data on the accuracy of belief prediction. %G eng %9 PhD thesis