SParSeFuL: Self-organizing sparse federated learning over spatially non-IID data

Federated learning has emerged as a promising solution for training machine learning models while preserving privacy in distributed environments. However, deploying federated learning in large-scale Internet of Things (IoT) systems presents significant challenges: data heterogeneity across spatially distributed devices, resource constraints of edge devices, and the lack of reliable centralized infrastructure. Classical federated learning approaches struggle with non-IID data and require a centralized server for model aggregation, which introduces bottlenecks and communication overhead. To address these limitations, we propose Sparse Proximity-based Self-Federated Learning (SParSeFuL), a decentralized, multi-model (namely, cluster-based) federated learning framework that forms federations at runtime of spatially proximate devices with similar data distributions. This enables personalized model training that respects local data patterns while maintaining global objectives. Crucially, SParSeFuL embeds model sparsification directly within the self-organization process; this tight coupling between model structure and federation topology shapes collaborative behavior and significantly reduces computational overhead and, as sparser models require fewer parameters to be exchanged, indirectly limits communication cost, making it highly suitable for resource-constrained IoT environments. To demonstrate the potential of SParSeFuL, we conduct an extensive experimental evaluation on two datasets widely used in computer vision: Extended MNIST and CIFAR100. Moreover, we provide a comparison against several state-of-the-art federated learning algorithms – namely FedAvg, FedProx, IFCA, and Scaffold – showing that SParSeFuL outperforms most baselines even at high sparsification levels.