Lightweight Meta-Learned RF Fingerprinting under Channel Imperfections for 6G Physical Layer Security

Chia-Hui Liu^*, Hao-Feng Liu
Department of Electronic Engineering, National Formosa University, Yunlin, Taiwan
* Corresponding Author: Chia-Hui Liu. Email: email
(This article belongs to the Special Issue: Artificial Intelligence for 6G Wireless Networks)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.077837

Received 17 December 2025; Accepted 11 February 2026; Published online 09 March 2026

Download PDF

Abstract

Artificial Intelligence (AI)-native sixth-generation (6G) wireless networks require data-efficient and channel-resilient physical-layer modeling techniques that learn stable device-specific representations under channel variations and hardware imperfections to support secure and reliable device-level authentication under highly dynamic environments. In such networks, massive device heterogeneity and time-varying channel conditions pose significant challenges, as reliable authentication must be achieved with limited labeled data and constrained edge resources. To address this challenge, this paper proposes an Artificial Intelligence (AI)-assisted few-shot physical-layer modeling framework for channel robust device identification, formulated within the paradigm of Specific Emitter Identification (SEI) based on radio frequency (RF) fingerprinting. The proposed framework explicitly formulates few-shot SEI as a channel-resilient physical-layer modeling problem by integrating a lightweight convolutional neural network and Transformer hybrid encoder with a dual-branch feature decoupling mechanism. Device specific RF fingerprints are separated from channel-dependent factors through orthogonality-constrained learning, which effectively suppresses channel-induced prototype drift and stabilizes metric geometry under channel variations. A meta-learned prototypical inference module is further employed under episodic few-shot training, enabling rapid adaptation to new devices and unseen channel conditions using only a small number of labeled samples. Experimental results on multiple real-world RF datasets, including ORACLE Wi-Fi transmitter measurements and civil aviation ADS-B broadcasts (DWi−Fi, DADS−B, and DADS−BDF17), demonstrate that the proposed method achieves identification accuracy ranging from 99.1% to 99.8% using only 10 labeled samples per device, while maintaining episode-level performance variance below 0.02. In addition, the proposed model contains approximately 1.45 × 10⁵ trainable parameters, making it suitable for deployment on resource-constrained edge devices. These results indicate that the proposed framework provides a concrete and scalable AI-driven solution for physical-layer security and device-level authentication in AI-native 6G wireless networks.