Adaptive momentum variance for attention-guided sparse adversarial attacks
作者：
Highlights：
• Our method considers two kinds of momentum variances, namely the forward momentum variance and the historical momentum variance, to adaptively stabilize the attack direction and escape from local optima.
• We refine the generated perturbation matrix to prevent the overfitting of the adversarial examples.
• We use the attention mechanism to perform the transfer-based sparse attack and assist in studying the relationship between the number of pixels attacked and attack performance.
摘要
•Our method considers two kinds of momentum variances, namely the forward momentum variance and the historical momentum variance, to adaptively stabilize the attack direction and escape from local optima.•We refine the generated perturbation matrix to prevent the overfitting of the adversarial examples.•We use the attention mechanism to perform the transfer-based sparse attack and assist in studying the relationship between the number of pixels attacked and attack performance.
论文关键词：Deep neural networks,Black-box adversarial attacks,Transferability,Momentum variances
论文评审过程：Received 15 May 2022, Revised 25 July 2022, Accepted 13 August 2022, Available online 21 August 2022, Version of Record 24 August 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108979
