Research On Lightweight Robust Live Face Detection Method

Face recognition technology is widely used in mobile payment,security access control,punch card check-in and other scenarios,but face recognition generally has the defect of being vulnerable to attack,people can use photos,videos,3D head models and other means to attack the face recognition system,so defending these attack means to make the face recognition system safe and reliable is an urgent problem to be solved.With the development of binocular vision sensor,it can bring rich features for face live detection,such as binocular structured light sensor and binocular stereo vision sensor can provide 3D stereo visual information,etc.Multimodal features can significantly improve the accuracy of face live detection methods,however,multimodal features need more parameters to learn,which is not conducive to the real-time system.Currently,convolutional neural networks are the main approach for live face detection,and this approach is not safe;there is usually a perturbation that is not easily detected by the naked eye,and this perturbation is enough to make the convolutional neural network make a wrong judgment.In order to make the binocular structured light sensor Intel Real Sense SR300 achieve the effect of real-time face live detection,this paper builds a lightweight face live detection network with multimodal feature fusion using depth features and RGB features based on the lightweight network Mobile Net V3,and quantifies the network with flow modules in the tail of the network,and solves the model for distinction of weight regions,and to enrich the model for feature extraction,this paper adds a bypass to the bottleneck layer of downsampling layer for improvement.Finally,a demonstration system is built and qualitative experiments are conducted on the public dataset CASIA-SURF and the CQNU-LN dataset produced in this paper,and the results show that the algorithm has good performance.In order to make the binocular stereo vision sensor Zed Mini achieve the effect of real-time face live detection,this paper designs a network architecture with three different fusion methods based on Feather Net,a lightweight network,and at the same time,in order to make the model focus on the intensity-level semantics as well as the gradient-level semantics,this paper uses central differential convolution to replace the convolution operation in the model,and then in this paper The experimental results show that the model satisfies the lightweight real-time criterion and possesses good results.Finally,for the application scenario of classroom check-in,a lightweight face live detection network is distilled using data-free distillation and combined with face recognition algorithm to build a classroom check-in APP.In order to defend against the adversarial perturbation information in the face live detection model,this paper proposes to add a noise reduction module to the module of the neural network,which plays the role of filtering the perturbation information to make the face live detection model more robust,and test the perturbation sample set generated by FGSM and C&W,and the experimental results show that the defense method has certain defense capability.