Feb, 2021
安全第一:通过对抗训练防范欺骗性对手
Better Safe Than Sorry: Preventing Delusive Adversaries with Adversarial Training
Lue Tao, Lei Feng, Jinfeng Yi, Sheng-Jun Huang, Songcan Chen
TL;DR本文证明了敌对训练可以作为防御欺骗攻击的可靠方法,并通过实验验证了其鲁棒性。敌对训练在自然环境中抵御欺骗攻击的机制是通过避免学习器过度依赖非鲁棒特征。