Research On Face Anti-spoofing Based On Deep Learning

With the development of information technology,various identity authentication technologies based on biometrics have been gradually commercialized,such as fingerprint unlocking,iris check-in and face payment.Face recognition technology has become a common biometrics recognition method because of its non-contact,obvious biometrics and easy access,which has brought huge social and economic benefits.At the same time,the fact that the face recognition system has been attacked also makes its security issue to be taken seriously.An attacker can spoof a face recognition system by forge a user’s facial information,such as printing face photo or playing face video.Therefore,it is of great significance to introduce the face anti-spoofing module into the face recognition system.Due to the diversity of face spoofing methods and the similarity between real faces and spoofing faces,it is difficult to distinguish them directly by the traditional features.The quality of the extracted spoof information will directly affects the result of the subsequent classification.How to represent the original image effectively and extract the appropriate features is a key issue in the field of face spoofing detection.The extensive application of deep learning in computer vision has demonstrated its effective ability of image feature extraction,making it becomes a research hotspot in the field of face spoofing detection.Therefore,this dissertation will further study the face spoofing detection technology based on deep learning to reduce the system error rate.This dissertation mainly focuses on the research of printing attack and video replay attack.In this dissertation,the method of face spoofing detection based on deep learning and related algorithms are analyzed.Based on the study of multi-dimensional information utilization,sample expansion and simulated label generation,three improved methods of face spoofing detection based on deep learning are proposed.The details are as follows:(1)We propose a two stream face anti-spoofing network combined with hybrid pooling.Firstly,optical flow is extracted from the dataset to represent the temporal dimension information of the dataset.Secondly,a hybrid pooling module combining space pyramid and global average is introduced to alleviate the network over-fitting by reducing the learning parameters.Finally,by weighting and fusing the results of the temporal dimension and the spatial dimension,improving the recognition effect.(2)We propose a face anti-spoofing network using local and depth information.Firstly,the dataset is divided into blocks to extract local information in the dataset,and the operation can expand the size of the dataset.Secondly,in order to guide the network to learn more effective features,depth information is introduced to assist the supervision of the network,and the structure similarity loss is combined to learn more robust depth information,which can extract the global information of the dataset.Finally,through the weighted fusion of local and global information,the recognition effect is improved.(3)We propose a face anti-spoofing network based on image conversion is proposed.Firstly,the simulated depth labels are generated from the 3D face model.Secondly,a high-quality generation model is obtained through the generative adversarial network,which is used to convert RGB images to depth images.Finally,use the generated features to train the classifier to achieve the purpose of making full use of the depth information.Due to the discrimination of depth information,the generalization ability of the model in unknown scenes is also improved.All the above methods are tested on public datasets such as Replay-Attack and CASIA-FASD.Different evaluation indexes are calculated to verify the effectiveness of proposed method.The comparison between the proposed method and the classical method shows that the proposed method is also competitive.