Federated Learning Scheduling Strategy Based On Particle Swarm Optimization Algorithm

As a popular AI technology,machine learning plays a vital role in providing intelligent services for the industry.The traditional machine learning method needs to collect data from the client and concentrate on the server to store and train the model,which will make the user’s private data have security and privacy problems.Federated learning is a new distributed machine learning method,which allows distributed medical institutions to upload model parameters to jointly train the model to avoid the serv er collecting sensitive data from the client.However,the federated learning server trains a common model through wireless communication with multiple clients,which increases the communication cost of the whole training process.In addition,due to diffe rent user preferences,the stored data is also heterogeneous,and the convergence time of model accuracy is also correspondingly improved,which may greatly reduce the performance of the model.In the field of health care,geographically distributed organi zations can upload model parameters to collaborate in training models,thus avoiding the server from collecting sensitive data from the client.This thesis studies the federated learning framework.The main work is summarized as follows:(1)In this thesis,we propose a dynamic balancing client data pre-processing method based on image enhancement.The local client sends its own label data amount to the central server,calculates the maximum data amount of each label of all clients through the server,and distributes it to the local client as the expanded quantity standard.The client expands and balances the existing data set through multiple image enhancement methods,so that the data amount of each label of the client is the same.This method can alleviate the impact of the decline of federated learning accuracy caused by the imbalance of data distribution.(2)This thesis proposes a federated learning optimization strategy based on particle swarm optimization algorithm,improves the process of federated l earning method,and introduces particle swarm optimization algorithm to optimize the federated learning training process.Before using local data for training,the client simulates each client model parameter as a particle,and optimizes the model to be tr ained according to the global optimal and individual historical optimal location information,It can obtain a better neural network model,improve the training efficiency of federated learning,and optimize the communication process of federated learning.(3)In this thesis,a federated learning method based on model weight divergence,update increment and training loss is proposed.By defining model weight divergence and update increment,the degree of non-independent and identical distribution of client d ata is identified.The larger the weight divergence and update increment of the client model,the higher the degree of non-independent and identical distribution of the client local data set.By selecting clients with lower non-IID degree to participate in training,and referring to the loss combination of client model training as the standard for selecting clients for training,select appropriate clients to participate in federated learning training at a higher frequency,instead of randomly selecting clients in traditional federated learning,and improve the effect of federated learning model.This thesis simulates and analyzes the publicly available chest disease data set including COVID-19,and compares it with the existing federated learning methods.The experimental results show that the method proposed in this thesis has better training performance in improving the accuracy of the model compared with the existing federated learning scheme,and reduces the impact of accuracy degradation caused by statis tical heterogeneity.