Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-uniform Attacks

Adversarial Training is proved to be an efficient method defend against adversarial examples, being one of the few defenses that withstand strong attacks. However, traditional defense mechanisms assume a uniform attack over examples according underlying data distribution, which apparently unrealistic as attacker could choose focus on more vulnerable examples. We present weighted minimax risk optimization defends non-uniform attacks, achieving robustness under perturbed test distributions. Our modified considers importance weights different and focuses adaptively harder are wrongly classified or at higher incorrectly. The designed allows training process learn through optimizing weights. experiments show our model significantly improves state-of-the-art accuracy attacks without significant drop