The article discusses GE-AdvGAN, a novel algorithm aimed at enhancing the transferability of adversarial samples while also improving the algorithm’s efficiency. This is achieved by optimizing the training process of the generator parameters. The algorithm uses a unique gradient editing (GE) mechanism and has proven effective in generating transferable samples on various models. By exploring frequency domain information to determine the gradient editing direction, GE-AdvGAN can generate highly transferable adversarial samples while minimizing execution time in comparison to other algorithms. The performance of GE-AdvGAN has been evaluated through large-scale experiments on different datasets, demonstrating the superiority of this algorithm.
Publication date: 12 Jan 2024
Project Page: https://github.com/LMBTough/GE-advGAN
Paper: https://arxiv.org/pdf/2401.06031
