Research On Hyperspectral-Based Face Spoofing Detection Method

With the continuous development of image acquisition and communication technologies,video chat,a convenient and practical way of communication,has become increasingly important in everyday life.And an important biometric feature of humans,the face,has been widely used in various fields such as device and software verification.However with the popularity of social media,face images have become more and more open and transparent,which makes them easily accessible to malicious attackers on the network.For example,attackers can conduct fraud,theft and other illegal activities by utilizing user’s faces through network acquisition and technical forgery.Therefore,as a critical verification step,face spoofing detection becomes increasingly important in scenarios such as video chat or user authentication.The existing methods have obvious defects that make them vulnerable to a replay attack launched by the virtual camera.This thesis argues that,for different types of attackers,it is far from enough to directly detect the collected images returned by the device.Instead,the challenge-response protocol can produce stronger artificial characteristics on the object to be detected,which can distinguish the adversarial intermediate media in a more safe and effective manner.Based on the study of facial skin reflection,we design a secure and effective challenge-response protocol that uses the special reflection characteristics generated by the screen light on the human face to detect face spoofing attack.We design adaptive face spoofing detection algorithms for two representative scenarios:In the video chatting scenario,this paper proposes deception detection by examining the spatiotemporal distribution of response sequences generated by illuminating the face with a sequence of challenges.To obtain response sequences from the facial videos illuminated by challenges,a response detection algorithm based on inter-band feature space distance using hyperspectral imaging is proposed,significantly enhancing subtle responses on the face.To further improve the robustness of the response detection algorithm,a band selection algorithm based on band scores is designed to select several bands with the lowest correlation and highest information content,effectively enhancing the sensitivity of inter-band feature space distance.In the software or device verification scenario,this paper proposes deception detection by combining the artificial response features generated from challenges with deep learning,training an end-to-end neural network.To effectively leverage the complementary features of the RGB space and hyperspectral space,a Multi-Space Feature Fusion Network(MSFnet)is proposed,which employs an adaptive feature fusion module to merge the output features of two feature extraction subnetworks,effectively enhancing the detection performance of different types of attacks.To reduce redundancy and noise between bands in hyperspectral images,a band fusion algorithm combining spatial and spectral information is introduced,which selects and merges several bands with the best classification performance,resulting in more discriminative hyperspectral features.Finally,extensive experiments are conducted on collected datasets,demonstrating high detection accuracy for genuine users and attackers under different conditions for both methods.This verifies the effectiveness and robustness of the two methods in various usage scenarios.