Batch Normalization (BatchNorm) is effective for improving the performance and accelerating the training of deep neural networks. However, it has also shown to be a cause of adversarial vulnerability, i.e., networks without it are more robust to adversarial attacks. In this paper, we investigate how BatchNorm causes this vulnerability and proposed new normalization that is robust to adversarial attacks. We first observe that adversarial images tend to shift the distribution of BatchNorm input, and this shift makes train-time estimated population statistics inaccurate. We hypothesize that these inaccurate statistics make models with BatchNorm more vulnerable to adversarial attacks. We prove our hypothesis by replacing train-time estimated statistics with statistics calculated from the inference-time batch. We found that the adversarial vulnerability of BatchNorm disappears if we use these statistics. However, without estimated batch statistics, we can not use BatchNorm in the practice if large batches of input are not available. To mitigate this, we propose Robust Normalization (RobustNorm); an adversarially robust version of BatchNorm. We experimentally show that models trained with RobustNorm perform better in adversarial settings while retaining all the benefits of BatchNorm. Code is available at \url{https://github.com/awaisrauf/RobustNorm}.

本研究证明BatchNorm在对抗攻击下表现脆弱的原因和提出了新的对抗攻击下稳健的正则化方法RobustNorm。实验证明使用RobustNorm的深度神经网络在对抗环境下表现更好。

走向一个对抗鲁棒性标准化方法