Design Of Efficient And Verifiable Aggregation Schemes For Federated Learning

With the rapid development of Artificial Intelligence,Deep Learning has also made major breakthroughs and is widely used in various industries.As a key driver of the development of artificial intelligence,data sources are shifting from large data centers to more and more mobile devices,but mobile device’s owners may be reluctant to share their local sensitive data with others due to privacy protection needs,thus bringing about the problem of ”data islands”.To solve this problem,model training needs to be used as much as possible to make use of multiple data,so a distributed federated learning framework was born.Federated learning involves many parties,complex management,and the potential for corruption and attack at each node,which threatens the global model.At the same time,although Federated Learning does not share local data directly,it still needs to exchange model parameters,and honest but curious servers can still infer other people’s information from it,causing privacy leaks.In addition,the aggregation server acts as a trusted third party in federated learning,but if the server is attacked or corrupts itself,it is possible to forge and tamper with the aggregation results.The existing scheme has performance problems such as inefficient and communication overhead,and lacks a verification mechanism for aggregating the results.In order to solve these problems,this article’s main work is as follows:An efficient federal learning security aggregation scheme is proposed.Most of the existing Federated Learning Security Aggregation schemes have large communication overhead,resulting in low feasibility of the solutions.In view of the above problems,we propose a new model based on the classic double-mask scheme under the regular graph setting.In view of the huge cost of secret sharing in existing schemes,the encryption and decryption overhead of homomorphic encryption algorithms,and other key issues,the balance between security and privacy is effectively realized.The framework adopts the federated learning training model based on the Top-K gradient selection scheme,combines the top-K gradient selection and secret sharing,compared with blindly uploading all the gradients,greatly reduces the calculation and communication overhead of the model,effectively improves the system performance,and improves the communication efficiency while ensuring user privacy and data security.Aiming at the lack of user verification mechanism for aggregation results and the single point of corruption of aggregation servers in federated learning,a verifiable secure aggregation scheme that can effectively verify the accuracy of server aggregation results is proposed.Using the Paillier homomorphic encryption algorithm,the message verification code with homomorphic nature is realized,and the user can compare and verify the correctness of the server aggregation result locally through the message verification code,so as to prevent problems such as invalid model training caused by server tampering with data.In this way,the confidentiality of the user gradient in the process of federated learning aggregation is guaranteed,the correctness and integrity of the aggregation results are ensured,and the accuracy of the model is improved while reducing communication and computing overhead.