Adaptive Meta-Learning for Personalized Federated Learning in Resource-Constrained IoT Environments

Abstract

Federated learning (FL) enables collaborative model training across decentralized devices without direct data sharing, proving particularly beneficial for Internet of Things (IoT) applications where data privacy and bandwidth limitations are paramount. However, the heterogeneity of IoT devices and their data distributions poses significant challenges for traditional FL algorithms. This paper proposes an adaptive meta-learning framework for personalized federated learning (AMFL-P) designed to address these challenges in resource-constrained IoT environments. AMFL-P leverages meta-learning to learn a personalized initialization and adaptation strategy for each device, allowing for faster convergence and improved performance even with limited local data. The framework dynamically adjusts the meta-learning process based on device resources and data characteristics. We present a detailed methodology, including a novel adaptive weighting scheme for meta-gradient aggregation. Experimental results on a simulated IoT sensor dataset demonstrate that AMFL-P outperforms traditional FL and existing personalized FL approaches in terms of accuracy, convergence speed, and resource utilization. The findings highlight the potential of adaptive meta-learning to enhance the effectiveness of federated learning in practical IoT deployments.