Revisiting single-step adversarial training for robustness and generalization

Open Webpage

Zhuorong Li, Daiwei Yu, Minghui Wu, Sixian Chan, Hongchuan Yu, Zhike Han

Published: 01 Jan 2024, Last Modified: 16 May 2025Pattern Recognit. 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•A correlation between catastrophic overfitting and local linearity in DNNs is identified.•A regularization is proposed to assure the validity of FGSM solution.•A thorough experimental study has been conducted.•Effectiveness shown via loss surface and decision boundary.