CS Events

Abstract: The Deep Neural Network (DNN) has achieved state-of-the-art (SOTA) results on many challenging tasks in computer vision, natural language processing and so on. With its wide adoption, the security of DNN becomes critical. One severe and important issue is the vulnerability of DNNs to backdoor attacks, where the adversary uses inputs stamped with triggers (e.g., a patch) to activate pre-planted malicious behaviors. In this talk, I will present our novel approaches for defending against backdoor attacks. For trigger reverse-engineering based defenses, we design a unified backdoor trigger inversion framework that can generalize to different types of triggers based on our novel formalization of the backdoor trigger. Additionally, we analyze the cause of DNN backdoors and conclude that linearity in DNN decision regions is the main reason. Based on our analysis, we propose a novel and general revised training framework that detects and fixes backdoors in DNN training.