Impact Analysis of Delayed Updates for Mobile Decentralized Federated Learning: A Delay-Tolerant Training Strategy for Real-World Edge Intelligence

doi:10.6688/JISE.20250541(3).0007

	Impact Analysis of Delayed Updates for Mobile Decentralized Federated Learning: A Delay-Tolerant Training Strategy for Real-World Edge Intelligence
	Zeng, Yong 1; Liu, Siyuan 4; Xu, Zhiwei 2,3; Tian, Jie 5
	2025-05-01
发表期刊	JOURNAL OF INFORMATION SCIENCE AND ENGINEERING
ISSN	1016-2364
卷号	41 期号:3 页码:627-644
摘要	Decentralized Federated learning is a distributed learning framework by learning a model with aggregated parameters among nearby participants, while keeping all the training data on the participants. Considering the various heterogeneous scenarios of mobile participants, the impact of transmission delay of updates during model training is non-negligible for data-intensive intelligent applications on mobile devices, e.g., intelligent medical services, automated driving vehicles, etc.. Although the latest solutions reuse the parameter updates for model parameter aggregating and approach a global model, there is no rational threshold for the delayed updates to guarantee model convergence. To address this problem, we analyze the impact of delayed updates for decentralized federated learning, and provide a theoretical bound for these updates based on augmented Lagrange function to achieve model convergence. Thereafter, we propose a novel decentralized federated learning scheme to enhance the corporate strategy of mobile computing devices. It releases the requirement for aggregating participants' updates within a specific time period, and provide a theoretical threshold for parameter updating delay. The latest versions for the delayed updates are reused to continue model training, in case the model parameters are not collected or updated within the theoretical threshold. Finally, we implement experiments on a real-world test bed, and demonstrate that delay-adaptive-DFL is more efficient than the latest baselines.
关键词	edge intelligence decentralized federated learning theoretical bound for delayed updates delay-tolerant decentralized training
DOI	10.6688/JISE.20250541(3).0007
收录类别	SCI
语种	英语
资助项目	State Key Laboratory of Networking and Switching Technology (Beijing University of Posts and Telecommunications)[SKLNST-2020-1-18]
WOS研究方向	Computer Science
WOS类目	Computer Science, Information Systems
WOS记录号	WOS:001488545700007
出版者	INST INFORMATION SCIENCE
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42386
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Xu, Zhiwei
作者单位	1.Sichuan Tengden Technol Co Ltd, Chengdu 611730, Sichuan, Peoples R China 2.Chinese Acad Sci, Inst Comp Technol, Beijing 100864, Peoples R China 3.Tianjin Univ Sci & Technol, Haihe Lab ITAI, Tianjin, Peoples R China 4.Inner Mongolia Univ Technol, Coll Data Sci & Applicat, Hohhot 010021, Mongolia, Peoples R China 5.New Jersey Inst Technol, Dept Comp Sci, Newark, NJ 07102 USA
推荐引用方式 GB/T 7714	Zeng, Yong,Liu, Siyuan,Xu, Zhiwei,et al. Impact Analysis of Delayed Updates for Mobile Decentralized Federated Learning: A Delay-Tolerant Training Strategy for Real-World Edge Intelligence[J]. JOURNAL OF INFORMATION SCIENCE AND ENGINEERING,2025,41(3):627-644.
APA	Zeng, Yong,Liu, Siyuan,Xu, Zhiwei,&Tian, Jie.(2025).Impact Analysis of Delayed Updates for Mobile Decentralized Federated Learning: A Delay-Tolerant Training Strategy for Real-World Edge Intelligence.JOURNAL OF INFORMATION SCIENCE AND ENGINEERING,41(3),627-644.
MLA	Zeng, Yong,et al."Impact Analysis of Delayed Updates for Mobile Decentralized Federated Learning: A Delay-Tolerant Training Strategy for Real-World Edge Intelligence".JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 41.3(2025):627-644.