Research On Performance Optimization Of Federated Learnin

Federated learning is a kind of architecture similar to distributed machine learning.Data storage and model training are performed by the client,and the global model is updated by exchanging model information with the server.The emergence of federated learning effectively resolves the contradiction between data privacy and traditional machine learning and becomes one of the rapidly developing branches in machine learning field.Federated learning is an emerging technology that can effectively protect personal information.Different from traditional centralized learning,federated learning can realize training and updating of global models while avoiding sharing private data.In the process of updating global model,the frequent exchange of model parameters between client and server takes up a lot of communication resources,which hinders the wide application of this technology.Based on federated learning,this paper optimizes the overall performance of federated learning from the following two aspects.(1)Aiming at the problem of communication consumption of federated learning,the federated learning is modeled as a two-objective optimization problem to reduce the communication consumption of the model without seriously reducing the accuracy of the global model.In this paper,the global model structure of sparse connection is constructed from the perspective of sparse global model,which reduces communication consumption in the communication process.Then,decomposition based multiobjective optimization algorithm is used to solve the optimization problem of model accuracy and communication consumption.Simulation experiments based on IID and non-IID data environment respectively verify that the algorithm can generate the neural network model of using in federated learning,and find the points with fewer connections and higher accuracy of the global model,thus significantly reducing the communication cost.(2)Privacy protection is an important feature of federal learning.Combining federated learning with differential privacy,secure multi-party computing and other technologies further strengthens the security of federated learning,but at the same time further increases the communication cost and computing pressure.In this paper,a federated evolutionary algorithm for medical EEG data sets is proposed by combining evolutionary algorithm with federated learning.The algorithm trains the global model without using gradients,reducing the data transfer from a single user to the server to a single fitness value.In this way,users do not need to upload the gradient information of the model,avoiding the problem of transmission data leakage,and greatly reducing the communication consumption in the process of communication.Experimental results show that this algorithm has certain advantages for data sharing in medical institutions.