Examining of Shallow Autoencoder on Black-box Attack against Face Recognition
Abstract
In this paper, we propose a Black-box Adversarial Examples (A.E.) attack that is effective for face recognition. Black-box A.E. for face recognition had multiple problems such as low probability of successful attack, limited attack targets, or large computational complexity which lead to impracticality in many real world scenarios. Therefore, we propose a more effective method of attacking face recognition system using Black-box A.E. by creating an attack substitute model suitable for face recognition based on the A.E. generation method of Huang et al. For evaluation, this method and the public dataset are used to attack arbitrary and specific people registered in the face recognition system which points out the possibility of a Black-box Adversarial Attack against face recognition system.