SOAR: Second-Order Adversarial Regularization

Avery Ma, Fartash Faghri, Nicolas Papernot +1

Adversarial training is a common approach to improving the robustness of deep neural networks against adversarial examples. In this work, we propose a novel regularization approach…