通信学报2026，Vol.47Issue(1)：13-26,14.DOI:10.11959/j.issn.1000−436x.2026012
基于置信度差异与熵最小化的跨时间鲁棒射频指纹识别方法
Robust RF fingerprint identification under temporal domain shifts via confidence discrepancy and entropy minimization
摘要
Abstract
In real-world wireless electromagnetic environments,radio frequency fingerprint(RFF)identification systems are inevitably affected by time-varying channel conditions and environmental changes,which lead to distribution mis-match between training and testing data.Such mismatch causes notable performance degradation across time-varying sce-narios,severely limiting the stability and generalization capability of RFF identification models in unknown environ-ments.To address this challenge,a cross-domain generalization-oriented test-time adaptation method for RFF identifica-tion was proposed.The proposed approach did not rely on labeled target-domain data and mitigated environment-induced domain shifts by adaptively updating the model during the testing phase.Firstly,an Inception-based RFF identification network was designed to enhance the robustness of fingerprint features under complex channel conditions by exploiting multi-scale feature representations.Secondly,considering the practical constraint that neither source-domain data nor target-domain labels were accessible during testing,a source-free and unsupervised test-time adaptation framework was developed,enabling the model to progressively adapt to the target-domain distribution.Experimental results on multiple public datasets demonstrated that the proposed method achieved superior identification performance compared with exist-ing approaches,while maintaining strong robustness and generalization capability in cross-time scenarios.These results validate the effectiveness of the proposed method for practical RFF identification deployment in complex environments.关键词
物理层安全/射频指纹识别/深度学习/测试时自适应/无源无监督域适应Key words
physical layer security/radio frequency fingerprint identification/deep learning/test-time adaptation/source-free unsupervised domain adaptation分类
信息技术与安全科学引用本文复制引用
张杰,王琴,尹悦,王禹,桂冠..基于置信度差异与熵最小化的跨时间鲁棒射频指纹识别方法[J].通信学报,2026,47(1):13-26,14.基金项目
国家自然科学基金资助项目(No.62471247,No.62401281,No.62472019)The National Natural Science Foundation of China(No.62471247,No.62401281,No.62472019) （No.62471247,No.62401281,No.62472019）
