Asymptotic Behavior of Adversarial Training in Binary Linear Classification.

Abstract

Adversarial training using empirical risk minimization (ERM) is the state-of-the-art method for defense against adversarial attacks, that is, against small additive adversarial perturbations applied to test data leading to misclassification. Despite being successful in practice, understanding the generalization properties of adversarial training in classification remains widely open. In this article, we take the first step in this direction by precisely characterizing the robustness of adversarial training in binary linear classification. Specifically, we consider the high-dimensional regime where the model dimension grows with the size of the training set at a constant ratio. Our results provide exact asymptotics for both standard and adversarial test errors under general -norm bounded perturbations ( ) in both discriminative binary models and generative Gaussian-mixture models with correlated features. We use our sharp error formulae to explain how the adversarial and standard errors depend upon the over-parameterization ratio, the data model, and the attack budget. Finally, by comparing with the robust Bayes estimator, our sharp asymptotics allow us to study the fundamental limits of adversarial training.