Master source code repo for "Batch Normalization is a Cause of Adversarial Vulnerability"
-
Section 1, Figure 1 and Appendix C: Adversarial Spheres
-
Section 3, Tables 1,2,3,4: SVHN / CIFAR-10 VGG experiments
-
Section 3, Figures 2, 3: MNIST MFT experiments. We reproduce the setup from Yang et al., ICLR 2019 and evaluate the trainability and robustness of fully-connected networks with 10-60 layers for mini-batch sizes 5-250, with and without batch norm.
-
Section 3, Table 4: CIFAR-10 and CIFAR-10-C experiments with unnormalized residual networks. Repo contains two pre-trained checkpoints for ResNet110 (with and without batch norm).
-
Section 4, Tables 6, 7, and Appendix B: Vulnerability and Input Dimension
Bibtex:
@article{galloway2019batch,
author = {Angus Galloway and Anna Golubeva and Thomas Tanay and Medhat Moussa and Graham Taylor},
title = {Batch Normalization is a Cause of Adversarial Vulnerability},
journal = {arXiv preprint arXiv:1905.02161},
year = 2019
}