Abstract
Federated Learning (FL) enables collaborative model training across distributed clients while preserving data privacy, but its deployment on resource-constrained devices is hindered by high communication overhead, inefficient energy usage, and poor convergence under non-IID data distributions. To address these challenges, we propose GRACE-FL: a Green Resource-Aware Communication-Efficient Federated Learning framework that explicitly incorporates device energy capacity into training. Each client adapts its learning rate, number of local epochs, and gradient quantization bit-width based on its available energy, allowing high-capacity devices to sustain more intensive training while low-capacity devices operate with lighter configurations. A novel energy-weighted aggregation strategy ensures that clients with greater energy resources contribute more significantly to the global model while maintaining fairness. Theoretically, we show that GRACE-FL achieves an unbiased gradient estimate and guarantees convergence at a rate of (Formula presented) under standard non-convex assumptions, even in the presence of quantization noise. Extensive experiments demonstrate that GRACE-FL consistently outperforms baseline methods, achieving 20–30% energy savings and up to 28% lower communication costs while preserving competitive or superior accuracy under both IID and non-IID settings. Finally, on-device validation with five Raspberry Pi 4B clients and fine-grained energy profiling confirms these gains in practice with SimpleCNN and ResNet-18, GRACE-FL delivers faster convergence, improved fairness, and substantial energy efficiency, paving the way for sustainable and practical FL in real-world edge deployments.
Recommended Citation
D. Thakur et al., "GRACE-FL: Green Resource-Aware Communication-Efficient Federated Learning," IEEE Transactions on Artificial Intelligence, Institute of Electrical and Electronics Engineers, Jan 2025.
The definitive version is available at https://doi.org/10.1109/TAI.2025.3636489
Department(s)
Computer Science
Publication Status
Early Access
Keywords and Phrases
adaptive gradient quantization; adaptive learning; energy Efficiency; Federated learning
International Standard Serial Number (ISSN)
2691-4581
Document Type
Article - Journal
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2025 Institute of Electrical and Electronics Engineers, All rights reserved.
Publication Date
01 Jan 2025
Comments
Università della Calabria, Grant PE00000013