Research On Federated Learning Algorithm Optimization Based On Model Distance And Similarity Matching

With the rapid development of Multi-Access Edge Computing and artificial intelligence research,Edge Intelligence that combines them together has become one of the most popular research fields in the current academic circles.Edge intelligence aims to use artificial intelligence technology to empower the edge side,so that data can be directly calculated and processed after being acquired by edge devices,thereby reducing the computing pressure on the central server and improving the overall efficiency.Federated learning is a crucial part of edge intelligence.Different from traditional machine learning or deep learning,federated learning can jointly train the final model under the condition that each computing node does not share their own local data,but only shares the model update in each training round.This computing method can protect the data privacy of nodes and reduce the problem of network bandwidth performance degradation caused by data transmission.However,due to the diverse types of edge devices and complex relationships,the data generated by different nodes may be None Independent and Identically Distributed(Non-IID),and there may be some malicious nodes in the edge devices that generate a large amount of redundant data to interfere with the federated learning process;In addition,due to the diversity of edge terminal requirements,there is a certain degree of correlation between the data or processing tasks of different terminal devices,and there may be certain permutation invariance in the parameters of multiple similar terminal models.These will affect the overall model accuracy and training convergence speed of federated learning,thereby reducing the overall collaborative efficiency.In order to solve the above problems,this paper conducts research on improving the collaboration efficiency of federated learning:(1)In order to solve the problems of data redundancy and non-IID,this paper proposes a federated learning edge model aggregation and filtering mechanism based on model distance(MDFedAvg),which uses the model distance to represent the degree of non-IID of each node.According to the degree of non-IID,different weights are assigned to each node,and the system will finally uses a weighted average method to aggregate the models of all nodes based on those weights,thereby reducing the impact of data non-IID distribution.In addition,the algorithm also proposes a node state detection mechanism,which can identify and exclude malicious redundant nodes according to the model distance,thereby ensuring the quality of all computing nodes.(2)In order to solve the problem of node data correlation,this paper proposes a federated learning mechanism based on network model sharing(NMSFed),which has two main steps.In step 1,NMSFed divides all computing nodes into multiple clusters based on a clustering algorithm;in step 2,NMSFed uses a neuron matching algorithm for each cluster to match and aggregate similar neurons in different clients together.Finally NMSFed generates an accurate and matching model for each cluster.Finally,the paper conducts simulation experiments on MDFedAvg and NMSFed algorithm to compares their performance with other federated learning algorithms,and finds that compared with other comparative algorithms,the accuracy and the convergence speed of MDFedAvg is least influenced by conditions when data is non-iid and redundant nodes exist,and NMSFed can achieve high accuracy within several communication rounds when the relevance of different nodes are high,thus proving that the algorithm proposed in the paper can better solve the problem of collaboration efficiency in federated learning.