A Two-Stage Explanation-Based Method Using Particle Swarm Optimization for Black-Box Adversarial Attack

Deep neural networks are vulnerable to adversarial attacks, where perturbations added to input images lead to incorrect predictions. Black-box adversarial attacks, which do not require access to the internal model, are more practical in real-world scenarios. While evolutionary algorithms have been employed for black-box attacks, their effectiveness is hindered by high-dimensional decision spaces. Therefore, we propose a two-stage explanation-based method using Particle Swarm Optimization (PSO) for black-box adversarial attacks. The first stage employs a multi-objective optimization approach to generate explanations by directly selecting feature subsets extracted through the Segment Anything Model (SAM). Subsequently, the second stage leverages PSO to generate perturbations based on the derived explanations. Experiments on ImageNet validate the method's effectiveness, demonstrating superior performance in solving the black-box adversarial attack problem.