Single Frame Face Anti-spoofing Detection Using Deep Learning

Because face recognition is more convenient and faster than traditional methods such as passwords,verification codes,and secret security questions,face recognition has been widely used in various practical scenarios,such as mobile phone unlocking,face-scanning payment and identity authentication,etc.It has become a part of our daily lives.At this time,some criminals used photos,videos,3D masks and other methods to attack the face recognition system,causing loss of privacy and property to users,and threatening the security of the face recognition system.In order to solve this problem,a face anti-spoof module c-an be added to the face recognition system,and the face recognition can be performed only after the face anti-spoof determines that it is a live face.The performance of face anti-spoof has an important impact on the security of the face recognition system.The main research content of this subject is how to use single-frame RGB face images for live detection,without the need for additional equipment such as depth cameras,and without the user to do some interactive actions(nodding,blinking,etc.)to improve user experience.The work of this paper is mainly divided into four parts:construct real multi-modal data set,lightweight face anti-spoof model,face anti-spoof network without nonliving feature,and face anti-spoof network based on auxiliary information.First,in order to verify the effect of the model in actual application before deployment,an Intel Real Sense D435 camera was used to capture a dataset of two modal images of Depth and RGB,which contained six attack methods and 20 subjects.Second,in order to better deploy the model on mobile devices,a lightweight face anti-spoof model was designed.Based on the lightweight model Feather Net,some improved methods such as high-pass filter,multi-mode feature fusion and OCCL loss function are added to improve the accuracy by 5.474% with almost no increase in the model complexity.Third,in order to prevent the model from extracting live-unrelated features for classification and leading to over-fitting,the corresponding network structure is designed from the two perspectives of extracting live-related features and removing live-unrelated features.The living-related features contained in different parts of the face image are similar,and more living-related features can be extracted by using the segmented face anti-spoof network.Use the auto-encoder to learn the internal model of the live face image,and subtract the output image from the input image of the auto-encoder to remove live-unrelated features.Fourth,in order to prove that the auxiliary information can be used to improve the performance of the model,two kinds of auxiliary information are used to improve the network respectively.The first is to use features such as illumination,distortion,and blur in the image quality evaluation to assist the face live detection model for classification;The second is to use the attention weight to assist the model for classification,and design an attention mechanism suita ble for the task of face living detection.