Variance-aware MGO-based intelligent client selection for communication- and energy-efficient federated learning in Non-IID Edge-IoT networks

The deployment of Federated Learning (FL) in Edge Internet of Things (Edge-IoT) networks faces significant challenges due to non-Independent and Identically Distributed (non-IID) data distributions, resource constraints, and communication bottlenecks. This research presents a communication-efficient FL framework, named FedMGO, that integrates Mountain Gazelle Optimization (MGO) for both local model optimization and intelligent client selection. Unlike conventional approaches, FedMGO employs MGO-based client selection that evaluates clients using composite fitness functions incorporating both accuracy and stability metrics based on performance variance, effectively identifying consistently reliable participants. The method employs dynamic exploration–exploitation balancing through adaptive parameter control and progressive annealing, ensuring robust convergence in heterogeneous environments. Comprehensive evaluations on CIFAR-10 under non-IID settings demonstrate that FedMGO achieves 71.58% accuracy, outperforming Federated Average (FedAvg) (65.50%) and Federated Particle Swarm Optimization (FedPSO) (68.10%). Remarkably, the proposed approach reduces the per-round communication cost by 80% compared to FedAvg by selecting only 2 out of 10 clients in each communication round. Owing to both reduced client participation and faster convergence (14 versus 28 communication rounds), FedMGO further achieves an overall communication overhead reduction of approximately 90% until convergence while maintaining higher accuracy and enhanced training stability in non-IID environments. Empirical execution time and energy consumption analysis further demonstrate that FedMGO achieves 57.5% energy savings and 41.3% faster total execution time compared to FedAvg, fully justifying the computational-communication trade-off. This accelerated performance stems from convergence in only 14 communication rounds, a 50% improvement over FedAvg (28 rounds) and 44% over FedPSO (25 rounds), where the marginal per-round computational overhead is substantially outweighed by the drastic reduction in global communication. The MGO-based variance-aware client selection proves particularly effective in maintaining model stability, with FedMGO exhibiting 8.2% higher consistency compared to conventional methods. Scalability analysis on federations of up to 100 clients confirms that FedMGO maintains consistent performance with only 1.23% accuracy degradation, demonstrating its practical deployability in large-scale Edge-IoT networks. These results establish FedMGO as a robust and efficient solution for practical Edge-IoT deployments, addressing critical challenges in resource-constrained FL environments through an optimal balance of computational intelligence and communication efficiency.