Face Anti-spoofing Based On Deep Learning

With the popularity of face recognition systems,in order to prevent them from the attacks of print photos and replay videos,face anti-spoofing is getting more and more crucial.However,the performance existing anti-spoofing methods is not convincing when dealing with video attacks or complex natural environments.The Convolutional Neural Network(CNN)is a powerful representation learning method and can learn the spatial features from a given image space.Long Short-Term Memory(LSTM)is a kind of recurrent neural network that learns the temporal features of videos.The Generative Adversarial Networks(GANs)consists of a generating model and a discriminant model,training with an Adversarial training procedure to obtain discriminative spatial features.Combining the characteristics of the three models,this thesis uses multi-features,spatial-temporal features and Latent Space features to perform face anti-spoofing on print photos and replay videos attacks.The main topics of this paper are listed as follows:(1)We reviewed the existing methods on face anti-spoofing algorithm and summarized the drawbacks of using single feature,which is merely effective under the single attack type.In addition,in order to distinguish the video attacks and real users,11 public available databases are analyzed.With the knowledge gained,we further experimented on 4 of them.(2)CNNs are used to learn multiple deep features from different cues of face images.Temporal features,color-based features and patch-based local features are integrated to perform face anti-spoofing.The performance is evaluated on three publicly available databases,which proves the effectiveness of the proposed method.(3)A CNN-LSTM structure is introduced to learn the high-level spatiotemporal features from the dynamic facial images.We extract deep features from both global and local face parts,i.e.eyes,nose and mouth,and then we fuse them for face anti-spoofing.This approach achieves better performance when evaluating on cross-database.(4)The latent representation,extracted from the middle layers of the GANs,can represent the space distribution of a given face image.Then,the latent representation from all the middle layers are fused and classified.The experimental results of several commonly used classification methods are compared.Finally,the method is evaluated in three databases.