A survey of gradient normalization adversarial attack methods
Jun Chen, Huang Qi-dong, Yunpeng Zhang · 2023
Recent research has found that deep neural networks are vulnerable and easily been attacked by adversarial samples. Improving the success rate of attacks against adversarial samples is a key issue in testing and improving the security and robustness of models, making adversarial samples a hot topic in current neural network research. The researchers found that the general adversarial attack problem can be transformed into an optimization problem to maximize the objective function. Currently, a important algorithms for solving such optimization problems are based on gradient methods, such as FGSM, I-FGSM, and other gradient symbol normalization methods. This article summarizes the research status of several typical adversarial attack algorithms and gradient-based normalization methods. The application of several gradient normalization methods in adversarial attacks and their algorithm convergence are prospected.