EA-AQF: energy aware adaptive quantization and freezing in federated learning
TL;DR
Energy-Aware Adaptive Quantization and Freezing (EA-AQF), a unified framework that co-optimizes communication and computation, is presented, a unified framework that co-optimizes communication and computation and maintains robust convergence in highly heterogeneous tasks.
Abstract
Federated Learning (FL) enables collaborative model training across decentralized edge devices while preserving data privacy. However, real-world deployments are severely hindered by high uplink communication overhead and heterogeneous on-device computational constraints. To address these joint bottlenecks, We present Energy-Aware Adaptive Quantization and Freezing (EA-AQF), a unified framework that co-optimizes communication and computation. EA-AQF introduces a tiered mapping policy that assigns stage-wise activation bitwidths based on a dynamic client resource indicator. To further reduce local training costs, we implement a deterministic, precision-conditioned stage-freezing policy that disables gradient updates for computationally expensive network stages during backpropagation. Additionally, weight quantization is applied to client updates to minimize the communication footprint. We evaluate EA-AQF across diverse architectures, including CNN-7, CNN-17, and ResNet-20 on CIFAR-10, CIFAR-100, and MNIST datasets, and LSTM and MLP on IMDB and HAR datasets, under varying statistical distributions. Experimental results demonstrate that EA-AQF achieves a consistent communication reduction of 72.9% to 74.8% and significant computation reduction up to 12.66\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}, which is 1.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document} higher than prior state-of-the-art methods, with a marginal accuracy trade-off typically within 2-3% of full-precision baselines. Quantitatively, our sustainability analysis reveals that EA-AQF achieves up to 65% reduction in energy consumption compared to state-of-the-art, while strictly limiting carbon emissions to as low as 0.0122 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm {kgCO_2e}$$\end{document} for resource-intensive tasks. Notably, in highly heterogeneous tasks, EA-AQF maintains robust convergence positioning it as a highly efficient solution for sustainable edge AI.