A Review of Deep Face Adversarial Sample Generation Techniques

With the wide application of deep learning in face recognition systems, it has achieved remarkable results in the fields of identity verification and security monitoring. However, it has been found that such systems are extremely vulnerable to adversarial samples, and the attacker only needs to add tiny, imperceptible perturbations to make the model output misclassified. The paper systematically reviews the definition, classification and typical generation methods of face adversarial samples, covering both physical and digital domain attacks. It further discusses the challenges of black-box and white-box attacks on detection systems and the effectiveness of mainstream defense means such as adversarial training and residual denoising networks. Finally, the deficiencies of current detection techniques in terms of generalization capability, deployment efficiency and physical attack defense are analyzed, and the future development direction of building a unified evaluation system and a multimodal robust detection framework is proposed. This research provides theoretical support and practical paths for the optimization of security protection and adversarial sample detection strategies for face recognition systems.