A black-box adversarial attack for poisoning clustering

作者：

Highlights：

• We propose a gradient-free optimization algorithm to effectively poison clustering algorithms.

• The algorithm possesses theoretically proven global convergence properties, meanwhile exceeding the performance of state-of-the-art algorithms.

摘要

•We propose a gradient-free optimization algorithm to effectively poison clustering algorithms.•The algorithm possesses theoretically proven global convergence properties, meanwhile exceeding the performance of state-of-the-art algorithms.

论文关键词：Adversarial learning,Unsupervised learning,Clustering,Robustness evaluation,Machine learning security

论文评审过程：Received 28 October 2020, Revised 23 August 2021, Accepted 6 September 2021, Available online 8 September 2021, Version of Record 17 September 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.108306