StealthMask: Highly stealthy adversarial attack on face recognition system

Abstract

Convolutional Neural Networks (CNNs) based on deep learning algorithms are widely used in real-world scenarios. However, these networks are vulnerable to adversarial examples-maliciously crafted inputs that can cause the model to make incorrect predictions. The existence of adversarial examples presents a significant challenge to the field of deep learning, with profound implications for various aspects of our lives. In face recognition technology, adversarial examples pose a substantial security risk. In this paper, we propose a novel method for generating adversarial patches designed to be worn as masks. The perturbed mask is crafted to deceive face recognition models, thereby highlighting the security vulnerabilities inherent in this technology. Our experimental results demonstrate that the mask generated by the proposed method effectively misleads the face recognition system, achieving high attack success rates while maintaining necessary stealthiness and transferability. Moreover, our method successfully attacks commercial face recognition systems and real-world access control systems, exposing the vulnerabilities of existing face recognition technologies in security-critical applications. Notably, compared to traditional methods, our proposed method emphasizes the stealthiness of the adversarial mask more than traditional methods. To account for physical-world factors, such as distortion, rotation, and deformations, we integrate a specifically designed loss function, thereby enhancing the method’s stability and reliability in practical scenarios.