Research On Efficient Communication And Multi-key Homomorphic Encryption Technology In Hierarchical Federated Learning Environment

At present,we are in the hot era of big data.How to effectively use big data is a research issue that enterprises,universities,governments and other institutions pay close attention to.Cloud computing enables multiple clients to centrally upload data to the cloud for joint training.However,with the disclosure of privacy issues and the government and people's attention to privacy and security issues,cloud technology is no longer applicable at the moment.In order to solve the problem of privacy protection,Google has proposed federated learning which has security performance guarantee.Each client can conduct joint training while keeping the data locally.It not only solves the problems of data islands among enterprises and devices,little high-quality data,unbalanced data distribution,slow communication speed and unstable communication connection.But it also keeps the data locally during the training process and reduces the risk of private information leakage.However,the current federated learning algorithm has low communication efficiency,high communication overhead,or server empty status,which affects the overall training efficiency.In terms of privacy protection,differential privacy technology is widely adopted.The addition of random noise to gradient parameters or model parameters can achieve a certain level of security.But it will affect the accuracy of model.The privacy protection design does not take into account the joint attack of multiple and internal malicious parties.On the premise that each client can effectively participate in the federated learning and maintain a high accuracy of model,reducing the communication overhead and defending against the joint attack of multiple and internal malicious parties is the main problem to be solved in this paper.In this paper,an in-depth study is conducted on the high communication overhead in the horizontal federated environment in federated learning.(1)This pepar analyzs the entire training and learning process of federated learning,and discussed the main influencing factors of communication overhead during the training process through the application of edge computing and asynchronous communication to federated learning.(2)This pepar designs a communication structure to join the edge server that changed the synchronous communication between the edge server and the cloud center server to asynchronous communication.Through using Convolutional Neural Networks(CNN)for training,it is proved that the accuracy of the model under this structure is improved compared with the two models of asynchronous communication model and edge computing.And the cost of communication is greatly reduced.(3)In order to solve the joint attack of multiple and internal malicious parties,a multi-key homomorphic encryption scheme was designed on the basis of the above communication structure.And the security analysis proved that the encryption scheme can resist the joint attack of multiple and internal malicious parties and depth gradient leak attacks.