Spoof Face Detection Via Semi-Supervised Adversarial Training.

摘要: Face spoofing causes severe security threats in face recognition systems. Previous anti-spoofing works focused on supervised techniques, typically with either binary or auxiliary supervision. Most of them suffer from limited robustness and generalization, especially the cross-dataset setting. In this paper, we propose a semi-supervised adversarial learning framework for spoof detection, which largely relaxes supervision condition. To capture underlying structure live faces data latent representation space, to train only, convolutional Encoder-Decoder network acting as Generator. Meanwhile, add second serving Discriminator. The generator discriminator are trained by competing each other while collaborating understand concept normal class(live faces). Since detection is video based (i.e., temporal information), intuitively take optical flow maps converted consecutive frames input. Our approach free faces, thus being robust general different types spoof, even unknown spoof. Extensive experiments intra- tests show that our method achieves better comparable results state-of-the-art techniques.