Oct, 2021
Learnable Adversarial Initialization 强化快速对抗训练
Boosting Fast Adversarial Training with Learnable Adversarial Initialization
Xiaojun Jia, Yong Zhang, Baoyuan Wu, Jue Wang, Xiaochun Cao
TL;DR本研究针对 Adversarial training 应用 Fast Gradient Sign Method (FGSM) 进行改进,提出了一种基于 generative network 和 target network 联合优化的 sample-dependent adversarial initialization 方案来提升模型的鲁棒性,并在四个基准数据库上进行了实验验证,结果表明该方法比现有的 Fast AT 方法更加优越。