FedAWM: Adaptive watermark allocation in non-IID federated learning

doi:10.1016/j.knosys.2025.114938

CSpace

	FedAWM: Adaptive watermark allocation in non-IID federated learning
	Sun, Jiahao 1,2; Yang, Xiaodong 1,2; Chen, Shubai 1,2; Qin, Xin 1,2; Zeng, Bixiao 1,2
	2026-01-15
发表期刊	KNOWLEDGE-BASED SYSTEMS
ISSN	0950-7051
卷号	332 页码:12
摘要	Black-box watermark is a widely adopted solution for protecting model copyrights in Federated Learning (FL), where additional watermark tasks are trained concurrently with the main tasks. However, a conflict between the gradients of the main and watermark tasks can lead to performance degradation, hindered convergence, and low watermark fidelity. This issue is further complicated by the non-independent and identically distributed (non-IID) data among FL clients, as the effectiveness of watermarking also varies across them if watermarks are injected uniformly. Inspired by the ability of black-box watermarks to be shared through FL aggregation, we proposed a novel method called Adaptive Watermark Allocation in non-IID Federated Learning (FedAWM) to optimize both main and watermark tasks. Instead of assigning watermark tasks uniformly, FedAWM evaluates each client's ability to accommodate watermark training and adjusts the embedding strength accordingly. The server periodically assesses clients' watermark performance and then allocates watermark samples proportionally and asymmetrically, prioritizing clients with a higher tolerance to task interference. This process ensures that capable clients embed stronger watermarks, while others receive reduced or no watermark assignments. Extensive experiments on four benchmark datasets demonstrate that FedAWM achieves high main-task accuracy and watermark detection fidelity, while also maintaining robustness under various attacks, including model extraction, distillation, and adversarial manipulations.
关键词	Federated learning Black-box watermark Non-independent and identically distributed (non-IID) Adaptive watermark allocation
DOI	10.1016/j.knosys.2025.114938
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Artificial Intelligence
WOS记录号	WOS:001631857700002
出版者	ELSEVIER
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42988
专题	中国科学院计算技术研究所
通讯作者	Yang, Xiaodong
作者单位	1.Chinese Acad Sci, Inst Comp Technol, Beijing, Peoples R China 2.Univ Chinese Acad Sci, Beijing, Peoples R China
推荐引用方式 GB/T 7714	Sun, Jiahao,Yang, Xiaodong,Chen, Shubai,et al. FedAWM: Adaptive watermark allocation in non-IID federated learning[J]. KNOWLEDGE-BASED SYSTEMS,2026,332:12.
APA	Sun, Jiahao,Yang, Xiaodong,Chen, Shubai,Qin, Xin,&Zeng, Bixiao.(2026).FedAWM: Adaptive watermark allocation in non-IID federated learning.KNOWLEDGE-BASED SYSTEMS,332,12.
MLA	Sun, Jiahao,et al."FedAWM: Adaptive watermark allocation in non-IID federated learning".KNOWLEDGE-BASED SYSTEMS 332(2026):12.