Research On Face Anti-spoofing Algorithm Based On Inter-class Relationship Learning

Face Anti-Spoofing is one of the most important research directions in the field of computer vision and biometric security.As the security guarantee in front of the face identity authentication technology,it plays a vital role in promoting the application of face identity recognition system,protecting users’ private property and guarding social and public security.However,most of the existing face anti-spoofing approaches regard this task as a simple binary classification problem between bona fides and spoofing faces.They do not pay attention to the inter-class relationship between different liveness types and ignore the differences among various attacks and the commonness between specific attacks and bona fides.This thesis aims to study the inter-class relationship among multiple liveness types,using the differences between various attack types and their commonness with bona fides to perform face anti-spoofing.We propose three efficient and accurate face anti-spoofing approaches.The main works are as follows:1)An inter-class commonness based face anti-spoofing approach is proposed.There are common features among different types.We use the inter-class commonness to train the network to extract more fine-grained classification features,thus improving the performance of the face anti-spoofing algorithm and alleviating the negative impact of introducing new attack types on existing models.Firstly,we propose a Common Feature Extraction Unit,which prior learns the common features of bona fides and two attack types from two perspectives of "Real + Mask" and "Real + Video" through a group-level classification strategy of "Real + 1 vs.Rest".Then,a Binary Classification Unit is proposed,which uses the common features obtained from the former module to classify the bona fides and spoofing faces.2)A multi-perspective feature learning based face anti-spoofing approach is proposed.Extracting features only from the perspective of commonness will make the model ignore the features unique to bona fide faces.Combining the common features with the unique features and extracting features from multiple perspectives,we can effectively improve the performance of the face anti-spoofing algorithm.We first propose a Multi-Perspective Features Extraction Unit,which uses both perspectives of "Live + 1 vs.Rest" and "Live vs.Spoof" for group-level classification to extract the common features between bona fides and 3D mask or video attack types,as well as the universal classification features unique to bona fides.Then,a Binary Classification Unit is proposed,combining the common features and universal features obtained from the former module to perform the final classification.3)An inter-class differences based face anti-spoofing approach is proposed.It is relatively inefficient to set all attack types into negative samples without further distinguishing the differences between them.We utilize a non-generative-adversarial disentangled network to explore spoofing features and separate user identity information and environmental background information irrelevant to attack detection.Considering the differences among various attack types,we use a fully convolutional structure with pixel-level cross-entropy loss to learn a multi-classification task and perform the face anti-spoofing task.