对抗训练可能会损害泛化

Jun, 2019

Adversarial Training Can Hurt Generalization

Aditi Raghunathan, Sang Michael Xie, Fanny Yang, John C. Duchi, Percy Liang

TL;DR本文研究了对抗训练在提高鲁棒精度（对抗方面）的同时又有可能降低标准精度（没有对抗方面）。通过构造凸学习问题，我们发现鲁棒精度和泛化能力之间存在基本的紧张关系，而利用未标记的数据进行鲁棒自我训练可以消除这种关系。

Abstract

While adversarial training can improve robust accuracy (against an adversary), it sometimes hurts standard accuracy (when there is no adve