Federated Learning Algorithm For Face Recognition

In recent years,massive data-driven deep convolutional neural networks(DCNNs)have greatly improved the accuracy of face recognition and verification.However,with the increasing attention to privacy in face recognition,face datasets are facing the risk of violating personal privacy.The current face recognition algorithms generally use the public face data set as the training set for end-to-end training on a server.We call this kind of method the traditional centralized training method.Considering the privacy problem,image data is often distributed everywhere,and the server will not have the right to access the massive face data in the real world.On the premise of protecting privacy,using decentralized face data for model training has important research and application value.At present,federated learning is the mainstream technology to solve model learning under privacy protection.In the traditional federated learning algorithm,the parameters of the whole network model are distributed and transmitted between many clients and servers as a whole.In the application of face recognition,the class feature vector of softmax layer can reconstruct the face image,and federated learning is easy to cause the problem of privacy disclosure.In addition,face recognition training needs many classes,so it needs more strict constraints to make its objective function similar to the traditional centralized training,so as to learn a more discriminative model.In view of the above two challenges,my main research work is as follows:Considering the problem of face information leakage caused by traditional federated learning,this paper first improves the traditional federated learning algorithm from the perspective of privacy protection.We modify the last fullyconnected layer of the convolutional neural network to ensure that each client holds a private fully-connected layer,so that clients have no right to access others’ class feature vectors,so as to realize privacy protection.Based on the traditional federated learning algorithm FedAvg,we changed the network optimization process to ensure that each client has a local objective function,and proposed FedPE(Federated Averaging with Private Embedding)algorithm.In this paper,CASIA-WebFace is used as the training set.Experiments on multiple face recognition benchmark datasets show that FedPE can initially converge and achieve good performance in face recognition and face verification.In order to further eliminate the difference between the optimization objective function of FedPE and the objective function of traditional centralized training,this paper proposes a federated learning algorithm FedGC(Federated Averaging with Gradient Correction)for face recognition based on gradient correction to reduce the difference of objective function.FedGC algorithm solves the problem of cross-client optimization that FedPE algorithm lacks.We begin to correct the gradient from the new perspective of back propagation,and introduce cross-client gradient to ensure that the update direction of the network conforms to the standard softmax loss function.We also make a theoretical analysis to illustrate the effectiveness and significance of our method.In this paper,CASIA-WebFace is used as the training set.Experiments on multiple face recognition benchmark data sets show that FedGC can surpass the performance of FedPE in face recognition and face verification,and even match the performance of traditional centralized training methods.