Reproduce “AdAUC: End-to-end Adversarial AUC Optimization Against Long-tail Problems” on OpenResearch

About this paper

Wenzheng Hou, Qianqian Xu, Zhiyong Yang, Shilong Bao, Yuan He, Qingming Huang

AdAUC introduces the first end-to-end adversarial training framework optimized for Area Under the ROC Curve (AUC) on imbalanced, long-tail datasets. This approach consistently outperforms accuracy-based adversarial training methods, demonstrating superior robustness and revealing that AUC optimization provides inherent resistance to adversarial perturbations.

Key insights

• Adversarial examples designed to minimize accuracy do not necessarily degrade AUC performance, highlighting a critical metric mismatch in traditional adversarial training.
• Optimizing directly for AUC, even in a natural training setting, can yield models with inherent robustness against adversarial perturbations.
• AdAUC methods learn more robust decision boundaries and maintain better class separability in score distributions and feature spaces under adversarial attacks.

Summary by alphaXiv.