Study On Algorithms Of Face Anti-spoofing Based On Deep Learning

Face anti-spoofing technology is a method to detect whether a face image is imitated by a real living human face.As an important security guarantee of face recognition technology,it becomes more and more important with the popularization of face recognition technology.With the development of this technology,the traditional spoofing attacks,such as photo attack and video attack,can no longer pass the detection of face recognition system.However,there are many new ways of spoofing attacks,such as using the eye image of legitimate users to block the eyes of illegal users.Traditional face anti-spoofing algorithms are based on some specific deception methods to classify live and fake human faces by artificial design features.In practice,the algorithms are unknown to spoofing attacks,so the generalization performances of traditional algorithms are poor.At present,there are few studies on unknown spoofing attacks,so this thesis studies the unknown spoofing attack detection algorithm based on deep learning,and improves it based on the application scenarios of face anti-spoofing technology.The work of this thesis is summarized as follows:1.An unknown spoofing attacks detection algorithm based on live face features is studied.The algorithm uses a convolutional neural network which combines attention mechanism and convolutional residual mechanism to extract live face feature maps from face images.The algorithm does not take the unknown spoofing features caused by the unknown spoofing attack as the basis for detection and classification,but the detection and classification are based on whether there is a complete live face feature map in the face image.At the same time,according to the features of live face features,face depth map is selected as the ground truth to optimize the neural network.The algorithm achieves Bona Fide Presentation Classification Error Rate of 7.1% and Average Classification Error Rat of 15.8% in the unknown spoofing attack scenarios simulated by SIW-M dataset,both of which are superior to the existing unknown spoofing attack detection algorithms.2.Due to the application requirements of face anti-spoofing technology,this thesis hopes that the algorithm will not easily classify fake faces into live faces,that is,to achieve lower Attack Presentation Classification Error Rate.It is necessary to enhance the sensitivity of the algorithm to the loss of live face features caused by spoofing attacks.By adding the image clustering network module,the face image routes were divided into clusters with similar spoofing features,and the live face feature networks corresponding to the four clusters were trained and optimized.Experiment shows,the improved unknown spoofing attacks detection algorithm sensitive to the loss of live face features achieves Attack Presentation Classification Error Rate of 8.9%,which achieves a performance improvement of 63.5% compared with the previous algorithm,and achieves Average Classification Error Rat of 14.5%,which is better than the existing unknown spoofing attacks detection algorithm.3.In order to verify the validity of the face anti-spoofing algorithm in real application scenarios,this thesis designs a face recognition system with the function of face anti-spoofing,and the effectiveness of the algorithm in the real situation is verified successfully.