Research on Attack and Defence of Face Recognition Based on Adversarial Learning

School of Intelligent Software and Engineering, Nanjing University, Suzhou, Jiangsu, 215004, China

221900068@smail.nju.edu.cn

Abstract

The rapid advancement of face recognition technology brings inevitable threats from various attacks, significantly reducing model accuracy and posing serious security risks. Enhancing defence mechanisms through algorithms is crucial to mitigate these impacts. However, these methods have limitations. New or complex attacks quickly diminish their effectiveness as attackers upgrade their techniques. Some defence techniques, while boosting model resilience, can adversely affect performance. This paper focuses on adversarial learning for attack and defence in face recognition, demonstrating the vulnerability of general models to attacks and giving a defence method that has some effect, while highlighting the limitations of current defences against complex attacks. Implementing adversarial learning reveals stark vulnerabilities: the constant tuning of parameters in the FGSM attack method reduces the model accuracy to 39.00%, illustrating its aggressiveness. In contrast, the adoption of PGD increases the sample prediction rate of the attacked prediction errors to 26.00%, demonstrating a viable defence strategy to bolster model robustness against such threats.