Safe and Robust Deep Learning - Gagandeep Singh
Workshop on Dependable and Secure Software Systems 2019
In this talk, I will introduce new automated reasoning methods for proving that deep neural networks satisfy a given specification (e.g., robustness against perturbations or other safety properties). These methods are based on abstract interpretation (specifically, new numerical approximations which scale to large networks) as well as combinations of these methods with mixed integer linear solvers. In the process, I will also briefly discuss how to apply these techniques not only for verification but also for training provably robust networks. The talk is based on results published at NIPS’18, POPL’19, and ICLR’19.