Learning deep models for face anti-spoofing by pixel-wise supervision with depth labels

Abstract

With the development of face recognition technology, vision-based face recognition systems are widely used. At the same time, various methods of attacking these face recognition systems have also begun to emerge. In this paper, we implement two types of anti-spoofing systems that detect such presentation attacks in a face recognition system using a deep learning network. The performance of the two systems was compared. A simple binary classifier using the entire face image and a depth information-based classifier that estimates depth information on a pixel basis is implemented. The performance of the implemented network was evaluated using the CelebA-Spoof database and the results of the two networks were compared.