Enhancing Adversarial Patch Effectiveness for Face Recognition Systems Using Sharpness-Aware Minimization

Facial recognition stands as a prevalent authen-tication solution within numerous biometric authentication applications. As a crucial alternative means, adversarial attacks serve to identify weaknesses in facial recognition systems and assess their robustness prior to deployment. Nonetheless, prevailing adversarial attacks encounter challenges such as large sample sizes and poor transferability. This study aims to develop a patch that reduces electronic domain similarity while maintaining high levels of generalization, based on adversarial mask. We introduce the Sharpness-Aware Minimization (SAM) optimization for adversarial examples, significantly enhancing their generalization across various facial recognition models. In contrast to perturbation methods focusing on specific indi-viduals, our technique offers a universal strategy that remains effective in real-world scenarios, demonstrating characteristics of generality, practicality, and transferability. Initially, we employ 3D facial modeling to simulate real-world conditions, followed by overlaying adversarial masks digitally. Through iterative optimization, the adversarial patch prioritizes minimizing the cosine similarity between genuine and adversarial embeddings. Our findings indicate a significant decrease in facial recognition accuracy, showcasing the potential of our approach in practical applications. The utilization of SAM not only enhances the robustness of patches against different models but also mitigates the risk of overfitting, representing a significant advancement in the field of adversarial machine learning.