Deep methods based on GAN for face-spoofing

Abstract

To prevent illegal access to users’ privacy by using face-spoofing, many researchers attempt to develop CNN models to identify it. However, only by working with high-quality face images, the CNN model can precisely report illegal accesses but is unreliable when the images are taken in bad conditions. To make up for the defect, this paper compares the performance of two kinds of more advanced neuron network models under the self-attention mechanism dealing with face-spoof issues. The first method is the Self-Attention GAN (SAGAN) model. Under the GAN framework, the SAGAN model introduces a self-attention mechanism to enable generator and discriminator to model the relationship between widely separated spatial regions. Based on this feature, SAGAN can be used to operate distant pixel points and then generate clear face images. The second research method is to apply the ViTGAN model to generate clear face images. Compared with developing the SAGAN model, general the ViTGAN model is a new approach, which can elegantly deal with the face images taken in the dark light conditions. This helps to CNN model cannot report face-spoofing issues with the input face images in the dark environment. To sum up, it is better to use the ViTGAN model to help to solve the face-spoofing issue.