Face Liveness Detection Based On Diffusion Model

Facial recognition technology is widely applied in various security systems,but these systems are vulnerable to spoofing attacks using photographs or videos,and there is a huge potential safety hazard.In order to prevent spoofing attacks,face liveness detection has emerged as the times require,and it has become an indispensable part of face recognition system.Face liveness detection is designed to determine whether the face image obtained by the system comes from the real legitimate user.Because of the rise of Internet social network,the acquisition of face data is very simple,and there are various kinds of ways to attack the system.The precision of fake face images is also improving.Therefore,face liveness detection is a challenging problem.In this thesis,we divide the task of face liveness detection into edge enhancement and feature extraction,and focus on how to extract the essential feature of live and fake face images to prevent various spoofing attacks.To model the relationship between face image sequence,a face liveness detection method based on Diffusion-based Kernel Matrix Model(DKMM)is proposed.The DK feature(Diffusion Kernel feature)is extracted from face images by DKMM.There are differences between live and fake face images in the image quality,so we apply anisotropic diffusion to enhance the edge of the image and blur the non-edge regions.This will not only preserve,but also sharpen the edges of the face images.In order to efficiently describe the difference between live and fake faces,the DK feature is extracted from the processed images.The feature can reflect the inner correlation of the face image sequence,and express the nonlinear relationship of the sequence in temporal dimension.The experiments show that our method can effectively prevent a variety of spoofing attacks even in complex conditions.To improve the liveness detection performance,a diffusion-based multi feature fusion method for face liveness detection is proposed.The DK features and the deep features are extracted from diffused face images.The DK features can reflect the inner nonlinear relationship of the face image sequence;The deep feature extracted by the pre-trained deep neural network model can reflect the inner hidden and abstract information of face images.Multiple kernel learning is used to fuse the extracted DK features and deep features to achieve better performance.The fused features are more discriminative and can provide a powerful basis for face liveness detection.The experiments show that our method has impressive performance and can distinguish live and fake faces effectively against spoofing attacks.To better utilize the characteristics of different face regions,a face liveness detection method based on multiple face regions is proposed.The face is divided into six regions: left eye,right eye,nose,left cheek,right cheek,mouth.The diffusion model is applied to each region to enhance the edge.We extract the feature based on co-occurrence of adjacent LBP to express the local characteristics of each region.The feature can efficiently model the relationship between the LBP features in spatial dimension.In order to focus on the face regions which have greater influence on liveness detection and achieve better performance,multiple kernel learning is applied to fuse the features of the six regions.Experiments show that the proposed method can utilize the characteristics of different regions.Owing to the fusion of local features of six regions,the method can effectively distinguish the live face from the fake ones.