Lightweight and Adaptive Federated Learning for Ultra-Low-Power IoT Devices
Main Article Content
Abstract
Federated Learning (FL) has emerged as a promising paradigm for decentralized model training while preserving data privacy. However, deploying FL on ultra-lowpower Internet of Things (IoT) devices remains challenging due to limited computational resources, communication constraints, and data heterogeneity. This paper proposes a lightweight and adaptive federated learning framework tailored for microcontroller-based IoT environments. The proposed method integrates quantization-aware training, gradient compression, and adaptive client selection to reduce communication overhead and energy consumption while maintaining model accuracy. Furthermore, a personalized aggregation strategy is incorporated to mitigate non-IID data distribution and domain shift issues. Experimental evaluation using a simulated IoT dataset demonstrates that the proposed method achieves up to 18% reduction in communication cost, 25% energy savings, and improved convergence stability compared to traditional FL approaches such as FedAvg and FedProx. The results validate the feasibility of deploying federated learning in resource-constrained environments, enabling scalable and privacy-preserving intelligent IoT systems.