Ultra-deep Neural Network for Face Anti-spoofing

Abstract

Face anti-spoofing is a hot research area in computer vision. With the progress of Deep Neural Networks (DNNs) in computer vision, some work has introduced neural networks into face anti-spoofing. However, the neural networks that most of the approaches use consist of only a few layers due to the limitation of training data. Inspired by the fact that deep efficiently trained neural networks are often possible to learn better representation than shallow networks. In this paper, we propose a fully data-driven ultra-deep model based on transfer learning. The model adopts a pre-trained deep residual network to learn highly discriminative features, and combines it with the Long Short-Term Memory (LSTM) units to discover long-range temporal relationships of from video frames for classification. We conduct extensive experiments on two most common benchmark datasets, namely, REPLAY-ATTACK and CASIA-FASD. Experimental results demonstrate that our ultra-deep network framework archives state-of-the-art performance.