Research On Federated Learning Method Based On Edge Computing

With the continuous development of mobile terminal technology,the number of devices connected to the Internet of Things continues to increase sharply,and the Internet of Things generates massive amounts of data.Using deep learning to analyze and process these data can dig out useful knowledge and provide users with more intelligent services.However,traditional deep learning applications need to transmit a large amount of user data to the cloud for processing,which will increase transmission costs and cause network congestion.Because edge computing can coordinate a large number of edge devices and process data at the edge of the network without transmitting to the cloud data center,it is a feasible solution to this problem.However,the scale of data collected by a single user is limited,and due to business competition and privacy protection,multiple users cannot share each other's data,making it difficult to train a satisfactory model.Federated learning is a distributed machine learning method that enables multiple participants to collaborate and train shared machine learning models without sharing data.Therefore,federated learning based on edge computing has become an effective method for processing massive Io T data,but there are still some challenges that need to be addressed.In order to better apply federated learning to the edge of the network,this thesis focuses on the following contents.First,to address the problem of insufficient communication efficiency of federated learning caused by the imbalance of user data distribution in edge computing systems,a method called Federated Learning with Data Distribution Weighted Aggregation(FLDWA)is proposed.The Hailinger distance is used to quantify the balance of the user data distribution of the participants,and the aggregation weight of each participant is adjusted accordingly,so that the algorithm can converge in fewer communication rounds or achieve the target accuracy rate to reduce communication costs.Second,in view of the problem that some users do not have the resource requirements to independently deploy complex deep learning tasks,which leads to the inability to participate in the training process of federated learning and the performance of federated learning is reduced,a deep neural network model deployment suitable for federated learning is proposed.Program.The deployment problem is modeled through the directed acyclic graph,and the deployment problem is simplified and decomposed using the idea of greedy algorithm,and finally the sub-problem is transformed into the minimum cut problem of the graph to be solved.This solution does not need to share data and deploys the model on multiple edge nodes,so that users of edge nodes can participate in federated learning,and the model training time is optimized.Corresponding simulation experiments are designed and implemented in this paper,and the proposed method is tested and evaluated to verify its feasibility and effectiveness.The experimental results show that under a variety of experimental settings,FLDWA can extract the information of each participant more efficiently,effectively alleviating the problem of insufficient communication efficiency.The proposed deep neural network deployment scheme also shows a faster model training speed in comparison with the benchmark scheme,and improves the accuracy of federated learning.