Research On Heterogeneous Federated Learning For Intelligent Internet Of Things

In recent years,with the continuous development of Internet of Things(IoT)technology and artificial intelligence(AI)technology,the degree of intelligence of Internet of Things devices has become stronger and stronger.This current intelligence is mainly through the use of IoT devices to aggregate large amounts of data and train models through centralized machine learning.However,this method of aggregating IoT device data to the cloud for model learning has network bottlenecks and end-user privacy leakage,which seriously hinders the promotion and application of the IoT.The federated learning for the IoT can realize shared model learning without data from the client,thereby effectively reducing the amount of data transmitted over the network and protecting the privacy of client data,which has important theoretical and applied research significance.Compared with traditional distributed machine learning,there are equipment heterogeneity,limited resources and data heterogeneity in the IoT environment,which will adversely affect the training efficiency,communication cost and model accuracy of federated learning.This thesis conducts research on heterogeneous federated learning based on the heterogeneity of the IoT.The specific work includes:(1)Aiming at the resource heterogeneity of terminal devices in the IoT environment,there are problems of low training efficiency and high communication cost in federated learning.This thesis proposes a data heterogeneous federated learning framework based on adaptive gradient uploading.In this framework,firstly,the relationship between model accuracy and client upload times in the asynchronous federated learning training process is analyzed,and a constraint relationship model between model accuracy and client upload times is established;The Euclidean distance of parameter gradients quantifies the value of client parameters and reduces the upload of low-value client parameters.Finally,a resource heterogeneous federated learning framework based on gradient adaptive upload is built.The experimental results show that compared with the existing methods,the method proposed in this thesis reduces the number of communications by 30%under the same accuracy of the model,which effectively reduces the cost of network communications.(2)Aiming at the problem of data heterogeneity in the Internet of Things environment,which leads to the decrease of the accuracy of the local model of federated learning,this thesis proposes a data-heterogeneous federated learning framework based on parameter matching and averaging.In this framework,firstly,using the permutation invariance of neural network to perform neuron similarity matching on the model parameters of the client,making the federated learning more sensitive to the data characteristics of different clients;The federated learning parameter fusion model;finally,the experimental verification is carried out on the WiFi signal gesture recognition dataset.The experimental results show that the proposed scheme achieves 90%new user prediction accuracy under the condition of a small number of users.