Research On Face Spoofing Detection

In recent years,with the improvement of computer vision technology,face recognition has been successfully applied to various authentication systems with high recognition accuracy and high robustness,such as the unlocking of smart phones and the mobile payment of Alipay.However,due to the advantage of face recognition,these face recognition systems can be easily fooled when the hackers use the biometric information of the user to attack these systems.This kind of face spoofing brings a huge security threat to the information and property of the users of these authentication systems,which seriously restricts the application of face recognition technology.For face anti-spoofing problem,many detection methods have been proposed to solve it.However,most of these detection methods classify the detection problem into a two-category classification problem,that is,all types of fake faces are divided into one class,and all types of real faces are divided into another class,and no independent analysis of different types of fake faces is performed.Unfortunately,putting all different types of fake faces together for analysis makes it difficult to find the common features that can distinguish these fake faces from the real faces.Based on this,a new face spoofing database containing multiple attack types is first constructed to solve the problem of limited training data.Then,by comparing different types of fake faces with real faces,the face spoofing detection task is decomposed into three sub-tasks: image attack detection,replayed video attack detection and 3D mask attack detection.Finally,through the solution of fusing different detection algorithms,a comprehensive detection algorithm capable of detecting multiple types of fake face attacks is proposed.The main work of the research on the sub-tasks and fusion algorithm is as follows:1.Aiming at the problem of image attack detection in face anti-spoofing,this dissertation proposes an image attack detection algorithm based on a CompactNet model.In the algorithm,by using the color difference between the real faces and the printed photos or the displayed images,a compact space generator is applied to learn a new compact space.Compared with conventional color spaces,the real and fake face samples in the compact space have the smallest intraclass distance and have the largest interclass distance.Apart from that,the pre-trained feature extractor is invoked to extract the feature description from compact space,which can eliminate the measurement error caused by external interference factors(such as illumination change,pixel misalignment,etc.).In training stage,considering the complexity distribution of face samples,a novel ”point-to-center”triplet training sample combination mechanism is introduced to make the triplet cost function degrade more stable.To prevent network overfitting problem,the gradient generated by the triplet cost function is only applied to the parameters of the compact space generator during the optimization process,while the pre-trained feature extractor parameters are fixed.Experiments show that the algorithm can effectively detect the printed photo or displayed image attacks.2.Aiming at the problem of replayed video attack detection in face anti-spoofing,this dissertation proposes a replayed video attack detection algorithm based on motion blur analysis.In the algorithm,a new 1D CNN network capable of describing the brightness changes is proposed by using the difference in brightness between the real face and the replayed video caused by the “delay effect” of the display screen.In addition,through the analysis of the motion patterns of real faces and replayed video attacks,it is found that there is a difference in motion blur between them.Therefore,a hand-crafted LSP feature that can describe the degree of motion blur is proposed.In detection process,the features from 1D CNN network and LSP are concatenated to detect replayed video attacks.Experiments show that the proposed algorithm has higher detection accuracy and stronger generalization ability than other detection algorithms.3.Aiming at the problem of 3D mask attack detection in face anti-spoofing,this dissertation proposes a 3D mask attack detection algorithm based on intrinsic image analysis.In the algorithm,it is found by the Lambertian theory that different materials have different reflectance maps.For face anti-spoofing,the real faces and fake faces are composed of different materials.Based on this,intrinsic image decomposition algorithm is first invoked to compute the reflectance maps of real faces and 3D face masks.Then,the information differences of the reflectance maps are analyzed to distinguish the real faces and face masks.On the other hand,due to the influence of heartbeat,the facial area of real faces will show dynamic changes of blood flow.Therefore,a three-dimensional network is proposed to extract the dynamic features caused by different materials and dynamic heartbeat.At end,during the detection process,the features extracted from the three-dimensional network are concatenated to complete the detection of 3D mask attack.Experiments show that the algorithm can effectively detect the 3D mask attack under the condition of constant material properties.4.For the problem of multi-type fake face spoofing,this dissertation proposes a comprehensive algorithm by combining different types of fake face attack detection algorithms.The algorithm for detecting multiple types of fake faces can solve the problem that the independent attack detection algorithm cannot work normally under the condition that the attribute of fake face attack in the actual application is unknown.Experiments show that the proposed multi-type face spoofing detection algorithm has good detection performance for different types of fake faces.