Adversarially Robust Learning via Entropic Regularization

In this paper we propose a new family of algorithms for training adversarially robust deep neural networks. We formulate a new loss function that uses an entropic regularization. Our loss function considers the contribution of adversarial samples which are drawn from a specially designed distribution that assigns high probability to points with high loss from the immediate neighborhood of training samples. Our data entropy guided SGD approach is designed to search for adversarially robust valleys of the loss landscape. We observe that our approach generalizes better in terms of classification accuracy robustness as compared to state of art approaches based on projected gradient descent.