Wasserstein Robust Classification Error: Surrogate Loss Misalignment and Distributional Attacks
The paper introduces Wasserstein Distributionally Robust Optimization (WDRO) to provide tighter theoretical robustness bounds for neural networks by considering worst-case expectations over a distribution ball. It develops novel Wasserstein Distributional Attacks (WDA) that construct adversarial examples by shifting samples along specific directions, leading to stronger attacks and revealing vulnerabilities missed by standard point-wise evaluation.