Gaussian class-conditional simplex loss for accurate, adversarially robust deep classifier training

Abstract

In this work, we present the Gaussian Class-Conditional Simplex (GCCS) loss: a novel approach for training deep robust multiclass classifiers that improves over the state-of-the-art in terms of classification accuracy and adversarial robustness, with little extra cost for network training. The proposed method learns a mapping of the input classes onto Gaussian target distributions in a latent space such that a hyperplane can be used as the optimal decision surface. Instead of maximizing the likelihood of target labels for individual samples, our loss function pushes the network to produce feature distributions yielding high inter-class separation and low intra-class separation. The mean values of the learned distributions are centered on the vertices of a simplex such that each class is at the same distance from every other class. We show that the regularization of the latent space based on our approach yields excellent classification accuracy. Moreover, GCCS provides improved robustness against adversarial perturbations, outperforming models trained with conventional adversarial training (AT). In particular, our model learns a decision space that minimizes the presence of short paths toward neighboring decision regions. We provide a comprehensive empirical evaluation that shows how GCCS outperforms state-of-the-art approaches over challenging datasets for targeted and untargeted gradient-based, as well as gradient-free adversarial attacks, both in terms of classification accuracy and adversarial robustness.