Perturbation-Optimized Black-Box Adversarial Attacks via Genetic Algorithm

Deep learning models often exhibit vulnerabilities to adversarial attacks, which has led to the development of various attack methods to evaluate model robustness and devise defense strategies. Currently, adversarial attacks are typically designed to target specific models, utilizing their own evaluation metrics. However, many black-box adversarial attack algorithms struggle to achieve comparable success rates to white-box attacks. In this chapter, we introduce comprehensive evaluation metrics for different adversarial attack methods. Additionally, we propose a novel perturbation-optimized black-box adversarial attack, referred to as our method, which aims to achieve attack performance comparable to white-box attacks. Our approach utilizes a genetic algorithm to evolve approximate optimal adversarial examples through a series of evolutionary operations such as initialization, selection, crossover, and mutation. A fitness function is specifically designed to evaluate the attack ability and perturbation control of each example individual. Moreover, a population diversity strategy is employed during the evolutionary process to ensure the acquisition of approximate optimal perturbations. We conduct extensive experiments to evaluate the performance of our method. The results, obtained through simulations and real-world applications, demonstrate that our method outperforms current state-of-the-art black-box attack methods in terms of both attack capability and perturbation control.