/robust-normalization

Investigation of how divisive normalization improves robustness against adversarial attacks.

Primary LanguageJupyter NotebookMIT LicenseMIT

robust normalization

In Miller et al. (2022), we observe that divisive normalization improves robustness against adversarial attacks. This repo contains code for reproducing this phenomenon and tools for understanding it from a geometric perspective, using the manifold analysis toolkit by SueYeon Chung.