A Security Mask Encryption Method In Local Multi-nodes Federated Learning Environment

The phenomenon of "data silos" in deep learning restricts the development of the industry.Google has proposed a federated learning method that enables many clients to jointly train models with trusted servers while the data is stored locally.The current research on federated learning mainly focuses on the issues of security and training efficiency.The research shows that the attacker can indirectly obtain the sensitive information of the client data set through the model parameters.Given the above problems,this thesis combines the local multi-node federated learning with the privacy protection mechanism based on secure multi-party computation.It proposes a federated learning training algorithm Mask-FL to improve the training efficiency and ensure the security of federated learning.The main work of this thesis is as follows:(1)This thesis implements a local multi-node federated learning framework using the data parallel training method and designs a data segmentation algorithm based on computing power.Each client of federated learning generates multiple local nodes to accelerate the training process according to its computing resource capability,which forms a local multi-node federated learning training structure.The nodes in the client have the advantages of mutual trust and straightforward communication,so the local multi-node federated learning framework has better training efficiency.The data segmentation algorithm performs data segmentation according to the computing capabilities of each node.It balances the training time of each node in the client,thereby reducing the impact of unbalanced data volume.(2)An adaptive mask encryption protocol based on secret sharing is proposed for the above-mentioned local multi-node federated learning framework.The reusable security adaptive parameter mask is obtained by secret sharing.The local nodes of the client add masks to the model to protect the security of the model parameters,and then the client sends it to the server for aggregation.The security analysis proves that the protocol can resist the threats from the client and the server when the honest and curious client and the central server participate in the attack.(3)The Mask-FL federated learning algorithm is proposed by combining the adaptive mask encryption protocol with the local multi-node federated learning framework.In the experiment,the Mask-FL algorithm is used to train the convolutional neural network model.By conducting independent experiments on the parameters of Mask-FL and comparing three different federated learning algorithms,the results prove that the Mask-FL proposed in this thesis can protect the The client’s data privacy,and can maintain a high accuracy rate.Mask-FL effectively reduces the global communication rounds and effectively improves the training speed of the federated learning model.The Mask-FL algorithm provides a reliable solution for the application of federated learning and extends the application scope of federated learning.The algorithm provides potential application value for more federated learning research.In the local multi-node federated learning scenario,Mask-FL can optimize the training effect of federated learning and provide an excellent data privacy protection effect,which is a good source for federated learning.