Tomo Miyazaki

Improving Adversarial Robustness via Activation Amplification and Attenuation

Taïga Gonçalves, Yongsong Huang, Tomo Miyazaki, Shinichiro Omachi

The existence of adversarial attacks is often attributed to the presence of non-robust features in neural networks. While prior defenses reduce their impact via pruning, masking, or feature recalibration, we instead propose to jointly learn to amplify and attenuate these signals…Read more