Semi-Supervised RF Fingerprinting: Contrastive Representation Alignment and Unknown-Class Discovery
IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 13063-13078
DOI: 10.1109/tccn.2026.3741936 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Ke Wang und 8 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben.
Abstract (Originalsprache)
Radio-frequency (RF) fingerprinting is a key physical-layer security technique for identifying wireless devices based on their unique hardware-induced signal characteristics. However, the efficacy of conventional deep learning models is often limited by two practical challenges: the scarcity of labeled training data and the presence of novel, unknown emitters not seen during training. To address these challenges, we propose the Semi-Supervised RF Fingerprinting: Contrastive Representation Alignment and Unknown-Class Discovery (SRCU), which integrates contrastive representation alignment, a dual cross-entropy objective, and a stability-driven consensus clustering mechanism. Our framework leverages contrastive learning to structure a discriminative embedding space, while a dual cross-entropy objective simultaneously refines decision boundaries for known classes and promotes the dispersion of unknown signals. A stability-guided consensus clustering module then operates on this structured space to discover novel emitter classes from the unlabeled data. We validate SRCU on public real-world long range (LoRa) and automatic dependent surveillance-broadcast (ADS-B) datasets, where it achieves state-of-the-art performance, especially under sparse labeling. With only 5% of signals labeled, SRCU achieves 69.74% accuracy on LoRa and 81.20% accuracy on ADS-B, outperforming the strongest semi-supervised baseline by 12.76 and 1.50 percentage points, respectively. As the label ratio increases to 90%, its accuracy reaches 96.83% on LoRa and 99.30% on ADS-B, approaching the performance of fully-supervised models. Furthermore, our method successfully discovers and clusters signals from unknown transmitters, demonstrating robust open-world performance. In summary, the proposed SRCU framework enables the joint identification of known emitters and discovery of multiple unknown classes within a single training process. By significantly improving RF fingerprinting accuracy in data-scarce scenarios, this work represents a critical step toward the robust, real-world deployment of Specific Emitter Identification (SEI) in dynamic wireless environments.
Zitieren
APA
Wang, K., Gao, Y., Wang, Z., Zeng, D., Huang, H., Wang, J., Liu, J., Qian, C., & Li, J. (2026). Semi-Supervised RF Fingerprinting: Contrastive Representation Alignment and Unknown-Class Discovery. IEEE Transactions on Cognitive Communications and Networking, 12, 13063-13078. https://doi.org/10.1109/tccn.2026.3741936
Vancouver
Wang K, Gao Y, Wang Z, Zeng D, Huang H, Wang J, et al. Semi-Supervised RF Fingerprinting: Contrastive Representation Alignment and Unknown-Class Discovery. IEEE Transactions on Cognitive Communications and Networking. 2026;12:13063-13078. doi:10.1109/tccn.2026.3741936
BibTeX
@article{Wang2026,
author = {Wang, Ke and Gao, Yiting and Wang, Zhiyin and Zeng, Deguo and Huang, Hao and Wang, Jiao and Liu, Jiaxu and Qian, Chen and Li, Jianqing},
title = {Semi-Supervised RF Fingerprinting: Contrastive Representation Alignment and Unknown-Class Discovery},
journal = {IEEE Transactions on Cognitive Communications and Networking},
year = {2026},
volume = {12},
pages = {13063--13078},
doi = {10.1109/tccn.2026.3741936}
}
EUVIMED