NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:926
Title:A New Defense Against Adversarial Images: Turning a Weakness into a Strength

This paper presents a new method for detecting adversarial attacks, based on an analysis of the robustness to Gaussian noise and the closeness to decision boundaries. Overall, a good paper addressing an important problem. The comparison to previous detection methods is reasonably extensive, while some discussion on the reverse cross-entropy learning-based method ( would be valuable.