Benchmarking and Optimizing Federated Learning with Hardware-related Metrics


Kai Pan (Institute of Computing Technology, Chinese Academy of Sciences), Yapeng Tian (University of Texas at Dallas), Yinhe Han (Institute of Computing Technology, Chinese Academy of Sciences), Yiming Gan (Institute of Computing Technology, Chinese Academy of Sciences)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Federated learning (FL) serves as an effective way of preserving data privacy at network training through offloading training tasks to different client hardware and aggregation. Real hardware-related metrics such as latency and energy consumption directly decide the performance and accuracy trade-off in federated learning frameworks, yet most FL optimizations do not use real hardware metrics. In this work, we propose to benchmark federated learning with real measured hardware metrics and optimize FL frameworks through tailoring training hyper-parameters before offloading tasks each round given hardware metrics. With two examples FedAvg and FedOpt, we demonstrate we can significantly save training energy by up to 97.2\% and training latency by up to 98.0\% while maintaining training accuracy.

Citation

@inproceedings{Pan_2024_BMVC,
author    = {Kai Pan and Yapeng Tian and Yinhe Han and Yiming Gan},
title     = {Benchmarking and Optimizing Federated Learning with Hardware-related Metrics},
booktitle = {35th British Machine Vision Conference 2024, {BMVC} 2024, Glasgow, UK, November 25-28, 2024},
publisher = {BMVA},
year      = {2024},
url       = {https://papers.bmvc2024.org/0369.pdf}
}


Copyright © 2024 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection