Communication-Efficient Federated Swarm Learning for SAG-CM on Heterogeneous Data

Résumé fourni par la source

Federated learning (FL) enables multiple clients to collaboratively train models without sharing private data, but it often struggles under non-IID distributions and tight edge budgets–particularly in systems such as space-air-ground collaborative monitoring (SAG-CM), where satellite, aerial and ground sensors generate heterogeneous, large-scale data across constrained networks. Inspired by the biological intelligence (BI) of gregarious organisms, we propose MSGWO-FSL, a novel edge-learning framework for SAG-CM/Federated systems. Our design integrates artificial intelligence (AI)-enabled stochastic gradient descent with bio-inspired multi-strategy grey wolf optimization: each client maintains a local ($\alpha,\beta,\delta$) leader-set to explore both weights and sparsity, and uploads only the binary mask of the$\alpha$model. The server aggregates the masks to preserve personalization by consolidating common structure while retaining client-specific sparsity, yielding sparse, faster on-device models with reduced communication overhead. Our convergence analysis theoretically demonstrates that MSGWO-FSL outperforms standard FL under mild conditions, and we derive a model-divergence upper bound that guides stable step-size choices under non-IID data. Extensive experiments on five public benchmarks and two public remote-sensing benchmarks demonstrate MSGWO-FSL surpasses nine state-of-the-art baselines in average accuracy while maintaining communication and computation efficiency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.