Soften to Defend: Towards Adversarial Robustness via Self-Guided Label Refinement

arXiv:2403.09101v1 Announce Type: new
Abstract: Adversarial training (AT) is currently one of the most effective ways to obtain the robustness of deep neural networks against adversarial attacks. However, most AT methods suffer from robust overfitting, i.e., a significant generalization gap in adversarial robustness between the training and testing curves. In this paper, we first identify a connection between robust overfitting and the excessive memorization of noisy labels in AT from a view of gradient norm. As such label noise is …

abstract adversarial adversarial attacks adversarial training arxiv attacks cs.cv cs.lg gap however networks neural networks overfitting robust robustness testing training type via