Batch Normalization is a Cause of Adversarial Vulnerability

Sep 25, 2019 Blind Submission readers: everyone Show Bibtex
  • TL;DR: Batch normalization reduces robustness at test-time to common corruptions and adversarial examples.
  • Abstract: Batch normalization (BN) is often used in an attempt to stabilize and accelerate training in deep neural networks. In many cases it indeed decreases the number of parameter updates required to achieve low training error. However, it also reduces robustness to small adversarial input perturbations and common corruptions by double-digit percentages, as we show on five standard datasets. Furthermore, we find that substituting weight decay for BN is sufficient to nullify a relationship between adversarial vulnerability and the input dimension. A recent mean-field analysis found that BN induces gradient explosion when used on multiple layers, but this cannot fully explain the vulnerability we observe, given that it occurs already for a single BN layer. We argue that the actual cause is the tilting of the decision boundary with respect to the nearest-centroid classifier along input dimensions of low variance. As a result, the constant introduced for numerical stability in the BN step acts as an important hyperparameter that can be tuned to recover some robustness at the cost of standard test accuracy. We explain this mechanism explicitly on a linear ``toy model and show in experiments that it still holds for nonlinear ``real-world models.
  • Keywords: batch normalization, adversarial examples, robustness
  • Original Pdf:  pdf
0 Replies

Loading