基于隐空间约束生成对抗网络的活体检测

Face anti-spoofing technology is critical to prevent face recognition systems from experiencing a security breach. Most of presentation attack detection (PAD) methods consider the task as a supervised binary classification problem. Many of these methods struggle to grasp adequate spoofing cues and generalize poorly. In this paper, we formulate the face anti-spoofing detection as an anomaly detection task to tackle the generalization issue. A novel deep network is proposed by using adversarial training under semi-supervised learning framework. The underlying structure of training data is captured in the image reconstruction space and can be further restricted in the space of latent representation in a discriminant manner, leading to a more robust spoof detector. In the test, the attacks are regarded as out-of-distributions samples that naturally exhibit a higher feature reconstruction error in the latent space than real samples in the dataset. Experiments show that our model is clearly superior over cutting-edge semi-supervised abnormal detectors and achieves state-of-the-art results on both intra- and inter-database testing.