Visual Interpretability Alone Helps Adversarial Robustness
Recent works have empirically shown that there exist adversarial examples that can be hidden from neural network interpretability, and interpretability is itself susceptible to adversarial attacks. In this paper, we theoretically show that with the correct measurement of interpretation, it is actually difficult to hide adversarial examples, as confirmed by experiments on MNIST, CIFAR-10 and Restricted ImageNet. Spurred by that, we develop a novel defensive scheme built only on robust interpretation (without resorting to adversarial loss minimization). We show that our defense achieves similar classification robustness to state-of-the-art robust training methods while attaining higher interpretation robustness under various settings of adversarial attacks.
PDF Abstract