QuantFL: Sustainable Federated Learning for Edge IoT via Pre-Trained Model Quantisation

Forum topic · 小凯 · 2026-03-19

Summary

QuantFL (arXiv:2503.13827) is a sustainable federated learning framework for IoT edge devices that leverages pre-trained model initialisation to enable aggressive, computationally lightweight quantisation. The authors—Charuka Herath, Yogachandran Rahulamathavan, and Varuna De Silva—show that pre-training naturally concentrates update statistics, allowing memory-efficient bucket quantisation without energy-intensive error-feedback mechanisms. On MNIST and CIFAR-100, QuantFL reduces total communication by about 40% with full-precision downlink, and achieves at least 80% bit reduction when quantising uplink or downlink. Under strict bandwidth budgets, it matches or exceeds uncompressed baselines, reaching 89.00% test accuracy on MNIST and 66.89% on CIFAR-100 with bucket quantisation using orders of magnitude fewer bits. The paper considers both uplink and downlink communication costs and includes ablation studies on quantisation levels and initialisation, positioning QuantFL as a practical green solution for scalable training on battery-constrained IoT networks.

Paper Overview

Field: Machine Learning
Authors: Charuka Herath, Yogachandran Rahulamathavan, Varuna De Silva
Published: 2025-03-18
arXiv: 2503.13827

Introduction

Federated Learning (FL) enables privacy-preserving intelligence on Internet of Things (IoT) devices, but its frequent uplink transmissions carry a high energy cost and a significant carbon footprint. Although pre-trained models are increasingly available on edge devices, their potential to reduce the energy overhead of fine-tuning has remained underexplored. QuantFL addresses this gap.

Key Contributions

Pre-trained initialisation enables aggressive quantisation: The authors demonstrate that pre-training naturally concentrates update statistics, allowing memory-efficient bucket quantisation without energy-intensive error-feedback mechanisms.
Communication savings: On MNIST and CIFAR-100, QuantFL reduces total communication by 40% with full-precision downlink (≈40% total bit reduction), and achieves ≥80% bit reduction when quantising the uplink or downlink.
Accuracy under strict bandwidth budgets: Bucket quantisation (BU) matches or exceeds uncompressed baselines, reaching 89.00% test accuracy on MNIST and 66.89% on CIFAR-100 with orders of magnitude fewer bits.
Comprehensive analysis: The framework accounts for both uplink and downlink costs, with ablation studies on quantisation levels and initialisation.

Conclusion

QuantFL offers a practical "green" solution for scalable federated training on battery-constrained IoT networks, combining pre-trained initialisation with lightweight quantisation to cut both communication volume and energy consumption.

*Original abstract (excerpt)*:

> Federated Learning (FL) enables privacy-preserving intelligence on Internet of Things (IoT) devices but incurs a significant carbon footprint due to the high energy cost of frequent uplink transmission. While pre-trained models are increasingly available on edge devices, their potential to reduce the energy overhead of fine-tuning remains underexplored. In this work, we propose QuantFL, a sustainable FL framework that leverages pre-trained initialisation to enable aggressive, computationally lightweight quantisation. We demonstrate that pre-training naturally concentrates update statistics, allowing us to use memory-efficient bucket quantisation without the energy-intensive overhead of complex error-feedback mechanisms. On MNIST and CIFAR-100, QuantFL reduces total communication by 40%...