Improving Adversarial Defense with Self-supervised Test-time Fine-tuning

Abstract

Although adversarial training and its variants currently constitute the most effective way to achieve robustness against adversarial attacks, their poor generalization limits their performance on the test samples. In this work, we propose to improve the generalization and robust accuracy of adversarially-trained networks via self-supervised test-time fine-tuning. To this end, we introduce a meta adversarial training method to find a good starting point for test-time fine-tuning. It incorporates the test-time fine-tuning procedure into the training phase and strengthens the correlation between the self-supervised and classification tasks. The extensive experiments on CIFAR10, STL10 and Tiny ImageNet using different self-supervised tasks show that our method consistently improves the robust accuracy under different attack strategies for both the white-box and black-box attacks.

Tasks

Reproductions