A Density-Adaptive DBSCAN-Based Federated Model Training Method for Distributed IoT Terminals

Abstract

Traditional federated learning is vulnerable to client noise, irregular local updates, and unequal device involvement because distributed IoT terminals provide diverse, sparse, and non-independent data streams. For dispersed IoT terminals, this research suggests a density-adaptive federated model training approach based on DBSCAN. Incorporate density-aware client grouping in federated aggregation, identify anomalous update patterns without a set number of clusters, and enhance the client clustering approach using adaptive neighborhood estimation. To lessen the impact of unstable terminals and preserve the helpful edge-side knowledge in the suggested strategy, combine local update statistics, gradient similarity, communication reliability, and data-density indicators. The suggested approach is evaluated experimentally against FedAvg, FedProx, clustered federated learning, Krum, trimmed-mean aggregation, and anomaly-filtered aggregation in the presence of non-independent data, communication fluctuation, terminal dropout, and malicious update disturbance. The findings show that in a heterogeneous IoT context, density-adaptive clustering shortened convergence by 28 communication rounds, increased general accuracy from 89.6% to 94.8%, and decreased aggregation deviation by 37.4%. The study offers a workable federated training paradigm for large-scale IoT networks under challenging communication contexts, non-IID data distribution, and terminal quality fluctuations.