Dissertation talk: Adaptive and Diverse Techniques for Generating Adversarial Examples

Presentation | December 12 | 11 a.m.-12 p.m. | 380 Soda Hall

 Warren He, Graduate Student Researcher, UC Berkeley

 Electrical Engineering and Computer Sciences (EECS)

In this talk, I present my research on adversarial examples in deep learning systems. Adversarial examples are inputs derived from natural examples, introducing small worst-case perturbations that cause systems to process them incorrectly. First, I present techniques for evaluating defenses against adversarial examples under an adaptive attacker threat model. Second, I discuss experiments with new adversarial example generation techniques that differ from previous methods in important ways.