Living Face Anti-Spoofing Recognition Based On Deep Learning

Living face anti-spoofing recognition is the security guarantee of face recognition system.At present,living face anti-spoofing technology is facing severe challenges,mainly including being sensitive to light changes and poor robustness of attack types.With the rapid development of deep learning,deep learning has become the mainstream method of living face anti-spoofing research because of its excellent feature expression ability,but it has the following three problems: first,most of the existing face anti-spoofing models of deep supervision network ignore the temporal motion information between multiple frames.Second,the representation of anti-spoofing clues by the network ignores the gradient information.Third,the network lacks selective attention to the features extracted at different levels.Aiming at the problems that the current depth information supervision network has poor effect on face depth estimation and is sensitive to illumination change,a living face antispoofing algorithm based on spatio-temporal depth information is proposed.The optical flow guided feature unit and convolution gated recurrent unit are used to extract the temporal motion information between multiple frames to help refine the estimated face depth map.The central differential convolution is introduced and the central differential gradient block is designed to replace the traditional convolution kernel to help the network aggregate the gradient level information and reduce the sensitivity of the model to illumination changes.At the same time,the depth contrast loss is introduced to compare the depth difference between adjacent pixels to constrain the network training.Aiming at the problems that the current depth information supervision network is difficult to find the key areas and poor robustness of attack types,a living face anti-spoofing algorithm based on depth and r PPG information is proposed.The attention mechanism is introduced and a multi-scale attention feature fusion module is designed to help the network pay selective attention in different scale feature extraction and guide the learning of more distinctive anti-spoofing clues.The recurrent neural network is introduced to supervise and estimate the r PPG signal of face video.The activity of face image is scored through the estimated face depth map and r PPG signal,so as to improve the robustness of the model to attack types.Comparative analysis of in library test on Oulu-NPU and Si W data sets and cross test on CASIA-FASD and Replay-Attack data sets.The experimental results show that the average classification error rate(ACER)index of the living face anti-spoofing algorithm based on spatio-temporal depth information on Oulu-NPU protocol 1 for environmental changes is 0.9%,which is 0.6% higher than the current mainstream methods,which shows that the algorithm can effectively improve the robustness of the model to the change of illumination intensity.The ACER index of the living face anti-spoofing algorithm based on depth and r PPG information on Si W protocol 3 for different attack types is 1.99%,which is4.46% higher than the current mainstream method,which shows that the algorithm can effectively improve the robustness of the model to attack types.