Research On Domain Generalization Algorithm For Face Anti-Spoofing

Face recognition technology has been widely applied in daily lives such as financial payment,mobile device unlocking,and security access control.However,face recognition is faced with security threats.Photos or videos containing faces could easily deceive face recognition systems.Therefore,face anti-spoofing technology has emerged.It has the ability to distinguish between real faces and liveness attacks and improve the security of face recognition systems.Although current face anti-spoofing methods perform well in known domains,there is still significant room for improvement when faced with unknown scenarios.To enhance the discriminative ability in cross-domain scenarios and improve the generalization of face anti-spoofing models,this paper makes the following three contributions:1)This paper proposes a domain generalization algorithm for face anti-spoofing based on dynamic convolution kernel generation.This algorithm could generate adaptive convolution kernels for each sample according to its domain information for further semantic feature extraction,which aims to enhance the feature extraction ability for samples from different domains.According to the fact that the intra-class similarity of real faces is higher than that of liveness attacks,an asymmetric center loss is designed to map all samples to a more reasonable feature space.Experimental results show that the proposed algorithm can achieve better performance in multiple cross-dataset experiments of face anti-spoofing,enhancing the generalization ability.2)This paper proposes a domain generalization algorithm for face anti-spoofing based on feature generation and hypothesis verification.To address the difficulty of mapping all face images to a shared feature space,this method models face anti-spoofing as a problem of distinguishing real faces from non-real faces and designs two feature generation networks to generate hypotheses for real faces and known live attacks,respectively,thus constructing distributions for real faces and known live attacks.Then,two hypothesis verification modules are used to determine whether the input face image is from the distribution of real faces or not.The algorithm performs well in both cross-dataset and cross-type experiments,demonstrating reliable generalization ability.3)This paper proposes a domain generalization algorithm for face anti-spoofing based on multi-domain mixup.This method disentangles the face features into liveness features and domain features.Liveness features contain liveness information,while domain features are irrelevant to liveness.Then,the domain features are mixed and enhanced to generate face images with diverse domains but the same liveness label,which help to further train the framework.This method not only improves the generalization ability,but also can be flexibly applied to semi-supervised and unsupervised domain adaptation scenarios.Therefore,the proposed algorithm has better scenario universality.