Research On Communication Optimization And Training Acceleration Technology Of Decentralized Federated Learning In Edge Computing Network

In recent years,benefited from the vigorous development of the Internet of Things(IoT)technology,massive amounts of data generated by intelligent terminal devices(such as smartphone and smart cameras)have greatly promoted the development of artificial intelligence.The traditional cloud computing requires collecting this massive data to cloud servers for processing and training.However,considering the privacy risks,high communication latency,and high bandwidth consumption caused by transmitting data to the cloud central server,edge computing has been proposed.Edge computing is a new computing model that integrates cloud,network,end,and intelligence.It takes cloud computing as its core,modern communication networks as its means,and massive smart terminals as its frontier.By optimizing resource allocation,it makes computing,storage,transmission,application,and other services more intelligent,with complementary advantages and deep collaborative resource scheduling capabilities.Edge computing technology can be applied to the integration of the Internet of Things and artificial intelligence to provide intelligent services with low latency,low bandwidth,and high security nearby.However,edge devices have characteristics such as limited communication resources,system heterogeneity,and data distribution heterogeneity,which seriously affect the training accuracy improvement of edge-side distributed model training(or federated learning,FL).To solve the above challenges,this thesis proposes a key technology research on communication optimization and training acceleration technology of decentralized federated learning in edge computing network,with the main research content as follows:To address the problems of limited communication resources on edge devices and data distribution heterogeneity among devices,this thesis proposes a decentralized federated learning mechanism based on data feature transmission and topology construction(DFL-DF).Specifically,edge devices exchange data features and feature extractors with their communication neighbors,rather than the models(or gradients).We construct a communication topology based on the similarity of data distribution between edge devices to match suitable neighbors for each device.By combining data feature transmission with communication topology construction,the proposed method reduces communication resource costs and improves testing accuracy under data distribution heterogeneity.Simulation experiments show that compared with the traditional model exchange-based method,DFL-DF can improve testing accuracy by 2.5%-9.1%under communication resource constraints,while reducing communication resource costs by 29.4%-82.06%when achieving similar performance.In order to further solve the problem of low training efficiency caused by heterogeneous devices in edge systems,this thesis further proposes a decentralized federated learning mechanism based on joint local update frequency and the number of data feature batch optimization(SDFL-HR),based on the above research.Specifically,we formalize the decentralized federated learning problem related to the local update frequency and the number of data feature batch,and propose a method based on greedy algorithm to determine the local update frequency and the number of data feature batch for each edge device.Simulation experiments show that compared with existing methods,when achieving the same target accuracy,SDFL-HR can reduce training time by 6.8%-75.6%,and improve model testing accuracy by about 12.2%under the same resource cost.