adversarial training (AT) is a popular method for training robust deep neural
networks (DNNs) against adversarial attacks. Yet, AT suffers from two
shortcomings: (i) the robustness of DNNs trained by AT is highly
本文使用神经切线核(NTK)对敌对训练(AT)过程和性质进行了深入研究,揭示了数据归一化对 AT 的影响以及批归一化层中无偏估计器的重要性,并通过实验探索了内核动力学和提出了更节省时间的 AT 方法,同时利用内核内的频谱特征解决了灾难性过拟合问题。据我们所知,这是首个利用内核动力学观察改进现有 AT 方法的研究。