Making Adversarial Examples More Transferable and Indistinguishable

• DOI: 10.48550/arxiv.2007.03838
• Publication Date: 2020-01-01
• Authors: Junhua Zou; Yexin Duan; Boyu Li; Wu Zhang; Pan Yu; Zhisong Pan
• Keywords: Transferability

Fast gradient sign attack series are popular methods that used to generate adversarial examples. However, most of the approaches based on fast cannot balance indistinguishability and transferability due limitations basic structure. To address this problem, we propose a method, called Adam Iterative Gradient Tanh Method (AI-FGTM), indistinguishable examples with high transferability. Besides, smaller kernels dynamic step size also applied for further increasing success rates. Extensive experiments an ImageNet-compatible dataset show our method generates more achieves higher rates without extra running time resource. Our best transfer-based NI-TI-DI-AITM can fool six classic defense models average rate 89.3% three advanced 82.7%, which than state-of-the-art gradient-based attacks. Additionally, reduce nearly 20% mean perturbation. We expect will serve as new baseline generating better indistinguishability.