Certified Distributional Robustness on Smoothed Classifiers

Abstract

The robustness of deep neural networks (DNNs) against adversarial example attacks has raised wide attention. For smoothed classifiers, we propose the worst-case adversarial loss over input distributions as a robustness certificate. Compared with previous certificates, our certificate better describes the empirical performance of the smoothed classifiers. By exploiting duality and the smoothness property, we provide an easy-to-compute upper bound as a surrogate for the certificate. We adopt a noisy adversarial learning procedure to minimize the surrogate loss to improve model robustness. We show that our training method provides a theoretically tighter bound over the distributional robust base classifiers. Experiments on a variety of datasets further demonstrate superior robustness performance of our method over the state-of-the-art certified or heuristic methods.