Proper Network Interpretability Helps Adversarial Robustness in Classification
This addresses the challenge of ensuring both interpretability and robustness in neural networks for AI safety and reliability, offering a novel defensive approach.
The paper tackles the problem of adversarial attacks that can hide from or manipulate neural network interpretability, showing theoretically and experimentally that it is difficult to prevent such attacks from causing interpretation discrepancies. It develops an interpretability-aware defense that promotes robust interpretation without adversarial loss minimization, achieving robust classification and interpretation, outperforming state-of-the-art adversarial training methods, especially against large perturbation attacks.
Recent works have empirically shown that there exist adversarial examples that can be hidden from neural network interpretability (namely, making network interpretation maps visually similar), or interpretability is itself susceptible to adversarial attacks. In this paper, we theoretically show that with a proper measurement of interpretation, it is actually difficult to prevent prediction-evasion adversarial attacks from causing interpretation discrepancy, as confirmed by experiments on MNIST, CIFAR-10 and Restricted ImageNet. Spurred by that, we develop an interpretability-aware defensive scheme built only on promoting robust interpretation (without the need for resorting to adversarial loss minimization). We show that our defense achieves both robust classification and robust interpretation, outperforming state-of-the-art adversarial training methods against attacks of large perturbation in particular.