Research On Face Anti-spoofing Methods Based On CapsNet

Recently,due to frequent crime cases of identity theft,the security of face authentication systems has been drawing broad attention.By using photos or videos of legal users,attackers can easily crack their target personal accounts,which seriously threatens the privacy and property of the users.Face anti-spoofing methods are the most important means of enhancing the security of systems.Although existing methods have effectively strengthened the face authentication systems,there are still some deficiencies as follows: On the one hand,most of the methods are based on deep neural networks,which are proved to be prone to serious decision errors due to imperceptible noise perturbations.It leads the methods to be extremely vulnerable to attacks of adversarial examples,i.e.they are lacking in adversarial robustness.On the other hand,facial data in real world are really complex,there are obvious differences in feature distribution between data domains corresponding to objects of different races or ages,but existing methods have poor generalization capability in cross-domain applications.To solve the above problems,this thesis mainly carries out the following work:(1)Aiming at the lack of adversarial robustness in existing face anti-spoofing methods,capsule network(CapsNet)is introduced to propose an adversarial robust face anti-spoofing model named FAS-CapsNet.The capsule structure and reconstruction mechanism of CapsNet are used to reduce the disturbance effect of adversarial samples,and the illumination features are extracted to increase distances between real and fake faces,which further enhanced the detection and defense capability of the model against face spoof attacks.The experimental results show that the model can not only defend general face presentation attacks,but also have significant defense against adversarial examples with higher concealment.(2)Aiming at the weakness of cross-domain generalization of existing methods and FAS-CapsNet,a face anti-spoofing model with enhanced generalization capability named FAS-CapsNet++ is proposed,which is based on the central difference convolution(CDC).The block consists of cascaded CDC layers is applied to improve the expression ability of the model for fine-grained key features,so as to improve the cross-domain generalization of the model.Experiments on the cross-ethnicity dataset CASIA-SURF-Ce FA proved that FAS-CapsNet++ has gained significant improvement of cross-domain generalization capability compared with FAS-CapsNet.