Research On Adaptive Federated Learning Framework And Algorithm For Edge Intelligence

Along with the rapid development of the Internet of Things technology in recent years,the era of the Internet of Everything has gradually become a reality within reach.In order to process the massive amount of data generated by IoT devices and meet the real-time and reliability requirements of new scenarios,edge intelligence emerges at the historic moment.Edge intelligence combines the advantages of artificial intelligence and edge computing and can make up for the shortcomings of the cloud computing model,and becomes a new research hotspot at present.While federated learning is one of the focuses of the research on the edge intelligence model framework.Federated learning can aggregate the local model trained by the client through the server,and achieve distributed learning while avoiding data leaving the device,thus can provide privacy protection features and make up for the security shortcomings of edge intelligence.However,when deployed in edge enviroment,federated learning still faces many problems.Due to the differences in computing and communication capabilities,edge devices present heterogeneous characteristics,making the aggregation of federated learning global models restricted by slow devices.And due to The differences in deployment scenarios,and the data generated by edge devices are non-independent and identically distributed,which greatly affects the performance of the federated learning model.At the same time,edge nodes also need a fair and reasonable incentive mechanism to ensure the motivation to participate in federated learning.In order to meet the above challenges,this paper analyzes the federated learning training process,and separately optemizes its client selection,global model aggregation and local model training,and designes an adaptive federated learning framework and algorithm for edge intelligence finally.Specifically,the main work and innovation are as below:(1)Aiming at the issue of fairness,this paper proposes an incentive mechanism based on the stackelberg game.Starting from the cloud-side collaboration framework,the federated learning participants are subdivided into server,client,and federated learning users,corresponding to the cloud center,the edge server and terminal nodes,which are included in a three-level and two-tier game system.And an incentive-driven algorithm FedID is proposed,which has realized that all participants maximize the utility to achieve the incentive effect.(2)Aiming at the problem of reduced accuracy of federated learning global models after aggregation caused by device heterogeneity,this paper proposes a window-based asynchronous federated learning algorithm FedWIN,which controls the way the federated learning server receives client updates by designing a time window,and introduces adaptive learning rate and obsolescence optimize the aggregation process,realizing the balance between synchronous aggregation and asynchronous aggregation,and improves the robustness of the global model to device heterogeneity.(3)Aiming at the problem of reduced client model accuracy caused by non-independent and identically distributed data,this paper proposes a personalized federated learning algorithm FedMMTL based on multi-task learning,which comprehensively uses model mixing and multi-task learning to build a personalized model for the federated learning client,and introduces it the adaptive parameters control the mixing degree of the model,and at the same time,the two stages of client training are controlled by cosine similarity,and finally improves the accuracy of the client model and reduces the degree of dispersion of accuracy.Through the above work,this paper realizes the optimization and improvement of the edge intelligence-oriented federated learning algorithm,and provides a new technical solution for the realization of edge intelligence deployment.