Hyperspherical class prototypes for adversarial robustness
作者：
Highlights：
• Adversarial robustness is addressed from a class boundary estimation perspective.
• The learning process is monitored by geometrically-inspired optimization criteria.
• Three optimization criteria for the hidden layer data activations are devised.
• The proposed method provides increased robustness to adversarial attacks.
• Without adverse effects to classification accuracy in clean data.
摘要
•Adversarial robustness is addressed from a class boundary estimation perspective.•The learning process is monitored by geometrically-inspired optimization criteria.•Three optimization criteria for the hidden layer data activations are devised.•The proposed method provides increased robustness to adversarial attacks.•Without adverse effects to classification accuracy in clean data.
论文关键词：Adversarial defense,Adversarial robustness,Hypersphere prototype loss,HCP loss
论文评审过程：Received 7 July 2021, Revised 25 November 2021, Accepted 7 January 2022, Available online 10 January 2022, Version of Record 15 January 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108527
