Internet Of Things Traffic Prediction Method Based On Horizontal Federated Learning

In recent years,the Internet of Things has developed rapidly,and the scale of traffic generated by Internet of Things devices is also gradually expanding.Network traffic prediction can realize the macro traffic trend perception.Existing network traffic prediction methods mainly focus on the improvement of prediction accuracy,but ignore the restrictions of practical application environment.Existing network traffic prediction methods mainly focus on the improvement of prediction accuracy,but ignore the restrictions of the practical application environment.Therefore,we propose a traffic prediction method applicable to the Internet of Things,focusing on studying and solving the three problems of "data silos",prediction efficiency and prediction accuracy of the Internet of Things traffic.We propose the Internet of Things traffic prediction method based on horizontal federated learning,including gradient compression algorithm for Internet of Things traffic prediction and traffic prediction method based on SMN3-CIFGA model.The main work is as follows:(1)Propose the framework of Internet of Things traffic prediction method based on horizontal federated learning.Design the overall workflow of the method and analyze some usage scenarios of the method.This method solves the "data silos" problem of Internet of Things traffic through the application of horizontal federated learning,uses the dynamic threshold gradient compression algorithm DTGCA to improve the efficiency of traffic prediction,and uses the SMN3-CIFGA model to improve the accuracy of traffic prediction.(2)A gradient compression algorithm for Internet of Things traffic prediction is proposed under the framework of the method,aiming at improving the efficiency of traffic prediction.DTGCA,the dynamic threshold gradient compression algorithm,is designed as a gradient aggregation algorithm within the framework.By controlling the compression range of gradient parameters through dynamic threshold,efficient distributed training of traffic classification model is realized.Finally,the experimental verification shows that it is feasible for the horizontal federated learning framework to conduct distributed training of the traffic classification model.The average training time of the gradient compression algorithm based on dynamic threshold proposed in this paper is reduced by about 6.21%,which effectively reduces the distributed training time on the premise of ensuring the classification accuracy.(3)A traffic prediction method based on SMN3-CIFGA model is proposed under the framework of the method,aiming at improving the accuracy of traffic prediction.The overall structure of the traffic prediction model is designed according to the problems faced by the actual deployment of the edge nodes.Firstly,the traffic classification model Simplified Mobile Net v3(SMN3)was designed to realize lightweight classification of Internet of Things data,and then the traffic prediction model Coupled Input and Forget Gate with Attention(CIFGA)was designed to complete the traffic prediction after classification,so as to realize the high accuracy of Internet of Things traffic prediction under the limited hardware environment.Through experiments,the average training time of SMN3 model is reduced by about 40%compared with the classic lightweight model MobileNet V3,and the classification accuracy reaches 95.96%.Compared with the traditional recurrent neural network model Long Short-Term Memory(LSTM),the overall training time of CIFGA model is reduced by 10.02%,and the prediction accuracy is improved to 97.61%,which proves that SMN3-CIFGA model can have a high prediction accuracy on the basis of ensuring the efficient operation of the model.