Federated Learning Optimization Algorithms For Data Heterogeneity Scenarios

Federated learning,a distributed machine learning framework,aims to break down data isolation while guaranteeing user privacy,and combine many participants to build enhanced federated models to achieve performance comparable to that of centrally trained models.This feature has allowed federated learning to be applied in numerous domains.However,in the real Io T environment,the heterogeneity of the data collected by each device can lead to the phenomenon of "weight divergence" during the training of the federated learning model,which slows down the convergence of the algorithm and reduces the inference accuracy of the global model.In data heterogeneity scenarios,random selection strategies may lead to slower convergence and degraded global model performance.To address this problem,we propose a client selection algorithm based on model performance and data distribution awareness.The algorithm evaluates the model performance by the loss value of the client and quantifies the data distribution of the client using the cosine similarity index to identify and select the local models participating in aggregation.Experimental results show that the algorithm can effectively improve the model accuracy and convergence speed of the federated learning algorithm in data heterogeneity scenarios.In data heterogeneity scenarios,existing aggregation algorithms do not take into account local model quality variability,resulting in degraded global model performance.To address this problem,we propose a dynamic model aggregation strategy based on multifactor awareness.The strategy uses the data scale and data quality of the client to quantify the contribution of the client and dynamically assigns corresponding aggregation weights to it,so as to reduce the impact of low-quality local models on the global model.Experimental results show that the strategy can effectively improve the model accuracy and convergence speed of the federated learning algorithm in data heterogeneity scenarios.Finally,this paper integrates the two algorithms into a new algorithm and verifies their effectiveness through comparative experiments.Experimental results show that the accuracy and convergence speed of the proposed algorithm in this paper outperform the other three baseline algorithms under different data heterogeneity scenarios.