Understanding deep face anti-spoofing: from the perspective of data

摘要

Face biometrics systems are increasingly used by many business applications, which can be vulnerable to malicious attacks, leading to serious consequences. How to effectively detect spoofing faces is a critical problem. Traditional methods rely on handcraft features to distinguish real faces from fraud ones, but it is difficult for feature descriptors to handle all attack variations. More recently, in order to overcome the limitation of traditional methods, newly emerging CNN-based approaches were proposed, most of which, if not all, carefully design different network architectures. To make CNN-related approaches effective, data and learning strategies are both indispensable. In this paper, instead of focusing on network design, we explore more from the perspective of data. We present that appropriate nonlinear adjustment and hair geometry can amplify the contrast between real faces and attacks. Given our exploration, a simple convolutional neural network can solve the face anti-spoofing problem under different attack scenarios and achieve state-of-the-art performance on well-known face anti-spoofing benchmarks.