Rethinking the validity of perturbation in single-step adversarial training

Yao Ge, Yun Li, Keji Han

Abstract: Highlights•Explore the impact of perturbation magnitude in single-step adversarial training.•Adjust perturbation magnitude to align with human visual perception.•Improve the accuracy-robustness trade-off in single-step adversarial training.•Address catastrophic and robust overfitting in single-step adversarial training.