Adversarial for Good – Defending Training Data Privacy with Adversarial Attack Wisdom
Shengyuan Pang, Yanjiao Chen, Jiangyi Deng, Jialin Wu, Yijie Bai, Wenyuan Xu · 2024
Machine learning models dazzle us with their incredible performance but may irritate us with data privacy issues. Various attacks have been proposed to peep into the sensitive training data of machine learning models, the mainstream ones being membership inference attacks and model inversion attacks. As a countermeasure, defense strategies have been devised. Nonetheless, a unified theoretical framework and evaluation testbed for training data privacy analysis is lacking. In this paper, we taxonomize representative attack methods regarding the attack objective, attack knowledge, and attack capability. As for the defense, we focus on the novel idea of turning adversarial attacks into privacy protection tools, hence the title adversarial for good. To provide an open-sourced integrated platform to evaluate different attacks and defenses. Our experiment results show that adopting adversarial example attacks and adversarial training for data privacy protection is compelling, which may motivate more efforts in transforming adversarial to good in the future.