Researchers have proposed a decentralized federated learning framework that replaces simple voting-based client selection with a strategy driven by particle swarm optimization, a population-based search process. In the proposed PSO-DFL framework, the swarm-inspired search process dynamically evaluates participating clients and selects those contributing most significantly to the global model, aiming to improve both model accuracy and training stability in systems where no central server coordinates training.
The motivation comes from a recognized weakness in existing decentralized federated learning approaches. Selecting high-quality clients for model aggregation is a critical factor in overall system performance, but current methods predominantly rely on simple voting mechanisms to filter clients.
Beyond the swarm-driven selection strategy, the framework introduces a fitness-based client filtering step designed to eliminate the influence of low-quality or anomalous models before aggregation. By screening out these contributions, the fitness-based filtering strategy is designed to accelerate the convergence of the learning process.
The reported experimental results focus on two standard benchmarks. On the CIFAR-10 dataset under non-IID settings, PSO-DFL achieved 76.2% accuracy along with the highest reported micro-average AUC of 0.932, outperforming the voting-based methods it was compared against. Robustness was tested separately on MNIST with 10% of clients acting maliciously; under that attack scenario, the framework maintained a final model accuracy of over 95%, which the authors present as evidence of strong robustness and generalization capability in decentralized heterogeneous environments.
Subject of Research: Dynamic client selection using particle swarm optimization in decentralized federated learning
Article Title: Decentralized federated learning with dynamic client selection based on particle swarm optimization
Article References: Lifang, C., Kailong, X., & Qi, D. (2026). Decentralized federated learning with dynamic client selection based on particle swarm optimization. Cluster Computing, 29(16), Article 916. https://doi.org/10.1007/s10586-026-06560-5
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06560-5
Keywords: federated learning, decentralized learning, particle swarm optimization, client selection, CIFAR-10, MNIST, non-IID data, robustness, model aggregation, machine learning, Decentralized, federated
News Source: Veronica Carney. (October 30, 2026). Particle Swarm Optimization Steers Client Selection in Decentralized Federated Learning. Scienmag.
