GAN based feature-level supportive method for improved adversarial attacks on face recognition

摘要

With the rapid development of deep neural networks (DNN), DNN-based face recognition technologies are also achieving great success and have been widely used in various applications which require high-accuracy and robustness. However, deep neural networks are known to be vulnerable to adversarial attacks, performed using images added with well-designed perturbations. To enhance security of DNN-based face recognition, we need to explore deeper the mechanisms of related technologies. In this paper, we propose a feature-level supportive method, BiasGAN, to improve the performance of universal adversarial attack methods. We insert this image to image translation preprocessor before conducting adversarial example generation. BiasGAN will search in the potential face space and can generate images with biased face feature, causing generated face images to be easier to perturb efficiently. Experimental results show that this approach improves both fooling ratio and average perturbation size significantly at different perturbation levels.