Face Anti-Spoofing Algorithm Based On Deep Learning

With the rapid popularization of face recognition and authentication systems,it is extremely important to prevent legitimate users' faces from being attacked by fake faces.As an important part to ensure the security of the face recognition and authentication system,face anti-spoofing detection technology has been widely used in important fields such as intelligent security,smart homes,and medical education and other important fields.Aiming at the identification of photo and video attack methods in the field of face anti-spoofing,this paper studies the algorithm from three different perspectives by enhancing the authenticity discrimination of the image and enhancing the spatiotemporal feature correlation of the image.The main innovative work of this paper is summarized as follows:(1)Aiming at the photo attack method in the face anti-spoofing task,combined with the local binary pattern(Local Binary Patterns,LBP)image local texture feature operator,TV-L1 optical flow approximation method(Optical flow method),optical strain feature map(Optical strain)and deep neural network design network.A new face detection algorithm named Deep optical strain feature map for face anti-spoong is proposed.The TV-L1 optical flow approximation method can extract the optical flow change information between adjacent frames,and the optical strain map is used to extract the small amount of movement in the optical flow map,and the optical strain feature map in the DOSF algorithm integrates the time series information between adjacent frames.Compared with the traditional deep learning-based detection method,it improves the face anti-spoofing detection performance in the photo attack identification scene.The DOSF algorithm obtained 99.79% and 98.2%recognition rates in the photo attack scenarios of two public data sets.The algorithm effectively improves the recognition accuracy of face anti-spoofing without user cooperation.(2)DOSF algorithm only extract the time sequence feature information between two adjacent frames,while video attack methods often require longer sequence dependencies to accurately distinguish true and false faces,and its recognition effect is not ideal in the video attack scene.This paper further proposes a face live detection algorithm based on temporal correlation dynamic network(Spatio-Temporal Correlation Dynamic Network model,STCD).STCD considers the feature changes between consecutive multiple frames of the input image as a set of discrete signals,uses the correlation information between feature changes to represent the image timing features,and finally combines deep learning to realize the discrimination of true and false faces.Experimental results show that the STCD algorithm can effectively detect photo and video attack methods.(3)In view of the problem that the correlation calculation method in STCD algorithm cannot learn from data and does not make full use of spatial information.Combining with Transformer,a dual-stream deep correlation network model(DSDC)that integrates spatio-temporal information with end-to-end optimization is proposed.DSDC uses the selfattention mechanism to calculate the correlation features of temporal changes between features,and Transformer is used to enhance the spatial domain feature representation,which can simultaneously mine the global temporal information and local spatial information of face data,greatly reducing the loss of information.Experiments verify that the dual-stream deep correlation network model can effectively detect false face images,which further improves the generalization ability of the model.