Research And Prototype Implementation Of Face Forgery Detection And Recognition Technology Based On Deep Learning

As the first barrier of financial services risk defense,authentication plays a key role in the consumer information and capital security.Among them,authentication technology based on face biometric has been widely used in security and payment fields.In recent years,with the innovation and development of deep learning technology,people can use convenient open source tools to edit face attributes at will,or even synthesize realistic face images,which are difficult to distinguish whether the image is real or fake by eyes.If criminals apply the face editing and generation technology in authentication services,it will cause serious security threats,which brings a great challenge to the field of financial risk control.Therefore,the research on face identification in this thesis mainly focuses on the two aspects of face forgery detection and face recognition.The specific work is as follows:1.A face recognition scheme based on cross-age data enhancement and feature compression is proposed,which is mainly improved from the perspective of data enhancement and feature compression.In terms of data enhancement,this thesis designed AgeGAN to generate face images of different age stages in the adversity-network simulation to expand face data and realize data enhancement.In the aspect of feature compression,this thesis uses principal component analysis to compress the extracted face features vectors.Finally,experiments on two feature extraction networks,VGG16 and ResNet18,proved that this scheme could effectively improve the accuracy of face recognition model while reducing the time consumption in the prediction stage.2.A face true or false discrimination scheme based on neural network characteristic heat map is proposed.In this thesis,three improvements are made on the basis of the MesoNet baseline model:first,after the face data is input into the VGG network,the appropriate characteristic heat map is obtained from the convolution layer of the VGG network.After the face data is shrunk to the dimension of the original sample,it is combined with the original sample as the input of the model.Secondly,the empty convolution,asymmetric convolution and global average pooling layer are introduced into the model to reduce the number of parameters of the network.Third,combining with the characteristics of the data set,Focal loss is used as the loss function of the model.Finally,we conducted experiments on public data sets and self-made face discrimination data sets,including face data generated by DeepFake,StyleGAN and FusVAE models,and finally verified that this scheme can effectively enhance the accuracy and performance of the network model.3.Based on the Python Flask lightweight framework,the functional design and prototype implementation of the proposed real and false face recognition algorithms are implemented.The Web system mainly provides functional modules such as identity authentication and information entry,and displays and applies face recognition and authentication algorithms.