A low-query black-box adversarial attack based on transferability

作者：

Highlights：

• A high success rate, low distortion with low queries number black-box adversarial attack is proposed.

• We randomly select a surrogate model to improve the success rate and efficiency of our attack.

• Dynamic coefficient is adopted to improve the efficiency of our attack.

摘要

•A high success rate, low distortion with low queries number black-box adversarial attack is proposed.•We randomly select a surrogate model to improve the success rate and efficiency of our attack.•Dynamic coefficient is adopted to improve the efficiency of our attack.

论文关键词：Adversarial sample,Black-box attack,Transferability,Neural network

论文评审过程：Received 17 December 2020, Revised 23 April 2021, Accepted 28 April 2021, Available online 14 May 2021, Version of Record 21 May 2021.

论文官网地址：https://doi.org/10.1016/j.knosys.2021.107102