Optimization Of Federated Learning Aggregation Algorithm And Parameter Sharing Mechanism For Edge Network

With the rapid development of Edge Computing(EC)and Internet of Things(IoT)technology,the type and quantity of datasets produced by digital equipment are facing explosive growth.Big data collection can improve the performance of machine learning model and achieve both economic and social benefits,but it also makes personal privacy protection face greater challenges.Federated learning is a new distributed machine learning paradigm,which was first proposed by Google to solve the problems of data privacy and information island caused by centralized training method in traditional machine learning algorithms.During the whole training process,the computing work involved in data(such as the update of local model)is carried out locally,and the data does not flow out.Thus,much equipment cooperates to train a global model becomes feasible under the condition of effectively ensuring users’ privacy.However,there are still two kinds of problems in the training of Federated learning model which is different from the traditional distributed learning in the edge network scenario:1)The participants of Federated learning are flexible,thus the participation of unreliable nodes may affect the training process.2)The local dataset held by different participants often obey NonIndependent Identically Distributions(Non-IID).The training method commonly used based on Federal Average(FedAvg)will be great effected during the training process.This paper focuses on the problems of unreliable nodes’participant and the Non-IID of terminals’datasets in the edge network federated learning scenario.Firstly,the paper proposes a federated aggregation optimization algorithm based on the importance of nodes(FedIM)based on FedAvg proposed by Google,aiming at the unreliability caused by the high dynamics of terminal nodes.The method adds the importance evaluation and improves the aggregation process of the common training process in Federal Learning.The evaluation process is based on node behavior rules and IPPO model,used to reduce the impact of local parameters submitted by unreliable nodes on the global model.The impact of different malicious nodes’participation condition on training tasks in Federated Learning is studied and tested on MNIST dataset.The simulation results show that the federated learning aggregation optimization algorithms based on both rule and IPPO model have better performances in model training effect and convergence speed than the FedAvg aggregation algorithm.Secondly,this paper proposes a parameter sharing mechanism based on local data enhancement and Local Outlier Factor(LOF)detection aiming at the Non-IID datasets in teraminal nodes,adapting to the training task with the participation of Non-IID nodes.In the local update phase,the data enhancement method based on proved synthetic minority oversampling technique(PSMOTE)is adopted to improve the impact of node dataset’s distribution on local update.In the aggregation stage,the local model weight is split,and anomaly detection is carried out to reduce the impact of node parameters of Non-IID on the global model.Simulation results demonstrate that the proposed method has significantly improved compared with the traditional training method FedAvg and can better reduce the impact of strong Non-IID nodes on the global model and improve the training effect.