Performance Analysis And Optimization Of Federated Learning In Edge-fog Computing Network

Federated learning is a new artificial intelligence technology,which obtains global model through cooperative training between user nodes.In federated learning,users do not need to upload local data,which can greatly reduce communication costs and protect users' privacy.With the continuous improvement of computing and storage capacity of edge devices,a large number of functions sink to the edge of the network,giving birth to edge-fog computing.In the edge-fog computing,model training can be carried out closer to the data generation,which reduces the latency,network traffic and improves the efficiency of the applications.The deployment of federated learning in edge-fog computing systems can not only take advantage of the computing and storage capabilities of edge devices,but also use artificial intelligence technology to support some data-driven and computation-intensive services.Because of the existing network deployment strategy in the edge-fog computing system,it is difficult to guarantee the model performance and training efficiency of federated learning.Therefore,in this paper,the performance analysis and optimization method of federated learning in edge-fog computing system is studied in detail.The main research work and innovation points can be summarized as follows:First,this paper presents a federated learning performance analysis and optimization method for edge-fog computing system.Because the computing and communication capabilities are limited in the edge-fog computing system,to ensure the federated learning performance in the edge-fog computing system,the federated learning framework is first deployed for the edge-fog computing scenario,so that user nodes can use the locally collected data to conduct model training without uploading data to the edge computing server.Through local model training and global federated aggregation,global models can be cooperatively generated in edge-fog computing systems.Secondly,in order to analyze the accuracy performance of federated learning in the edge-fog computing system,the upper bound of the federated learning model accuracy loss is theoretically analyzed based on the statistical theory,and the concrete upper bound expression is derived.Finally,in order to better balance the model accuracy and performance of the edge-fog computing system with the consumption of computation and communication resources during the training process,a joint optimization problem is first proposed.Then,a joint optimization algorithm is designed which could balance the model accuracy performance and resource consumption of federated learning.The simulation results verified that the proposed optimization algorithm can achieve significant performance gain and reduce resource consumption compared with the unoptimized federated learning mode.Second,this paper proposes a learning paradigm of federated learning with joint opportunistic centralized learning,and studies the performance analysis and optimization methods under this learning paradigm.In order to make better use of the data collected from user nodes and the large amount of computing resources in edge computing server,first of all,this paper proposes a learning paradigm of federated learning model with joint opportunistic centralized learning.This learning paradigm uploads data for centralized learning when local model is training in federated learning,and ingeniously combines centralized learning with federated learning.Secondly,the theoretical performance of the proposed new learning paradigm is analyzed in detail,and a theoretically explicit upper bound expression of model accuracy is derived.Finally,in order to balance the assumption of computing and communication resources and the model accuracy performance,the joint optimization problem of model accuracy,latency and energy consumption is established,and the corresponding joint optimization algorithm is designed to solve the problem.Simulation results show that for federated learning with joint opportunistic centralized learning,compared with single learning mode,the proposed algorithm can effectively improve the model accuracy performance,and save computing and communication resources.