Adversarial Attack Generation Based on Min-Max Optimization

AI summaryThe researchers propose a general framework for adversarial attack generation based on min-max optimization, reformulating the minimization of the worst-case attack loss as a min-max problem that maximizes domain weights over multiple attack domains. The framework is applied uniformly to three types of attack generation: attacking model ensembles, designing universal perturbations across multiple inputs, and constructing attacks robust to data transformations. The authors state that experiments show the method significantly improves attack effectiveness over existing heuristic strategies and also yields robustness improvements over state-of-the-art defenses trained against multiple perturbation types; the adaptive domain weights learned from the framework can also be used to explain the attack difficulty of each domain. Code is open-sourced at https://github.com/wangjksjtu/minmax-adv。