Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off
TL;DR
- Robustness improves without sacrificing discrimination: AR-AT learns adversarially invariant yet discriminative representations.
- Asymmetric regularization reduces gradient conflict: Uses stop-gradient and a predictor to improve optimization.
- Split BatchNorm separates input distributions: Handles the mismatch between clean and adversarial representations.
Type
Publication
ICLR 2025
Presentation Slides
