Bowei Xi, A Game Theoretic Approach for Adversarial Machine Learning -- When Big Data Meets Cyber Security

Nowadays more and more data are gathered for detecting andpreventing cyber attacks. Unique to the cyber securityapplications, learning models face active adversaries that try todeceive learning models and avoid being detected. Hence futuredatasets and the training data no longer follow the samedistribution. The existence of such adversarial samplesmotivates the development of robust and resilient adversariallearning techniques. Game theory offers a suitable framework tomodel the conflict between adversaries and defender. We develop agame theoretic framework to model the sequential actions of theadversaries and the defender, allowing players to maximize theirown utilities. For supervised learning tasks, our adversarialsupport vector machine has a conservative decision boundary,whereas our robust deep neural network plays a random strategyinspired by the mixed equilibrium strategy. One the other hand,in real practice, labeling the data instances often requirescostly and time-consuming human expertise and becomes asignificant bottleneck. We develop a novel grid based adversarialclustering algorithm, to understand adversaries' behavior from alarge number of unlabeled instances. Our adversarial clusteringalgorithm is able to identify the normal regions inside mixedclusters, and to draw defensive walls around the center of the normalobjects utilizing game theoretic ideas. Our algorithm alsoidentifies sub-clusters of adversarial samples and the overlapping areaswithin mixed clusters, and identify outliers which may bepotential anomalies.