Generalizable and Interpretable RF Fingerprinting with Shapelet-Enhanced Large Language Models

TL;DR

This work tackles RF fingerprinting under domain shift by integrating pre-trained large language models with learnable 2D shapelets to achieve both generalization and intrinsic interpretability. A CNN-based input embedding maps I/Q signals into the LLM space, while a shapelet network captures discriminative local patterns; the two representations are fused for classification and few-shot prototype-based inference. The method demonstrates strong cross-domain and cross-dataset generalization, with notable few-shot gains (e.g., up to 1-shot 73.76% on ORACLE) and interpretable explanations via learned 2D shapelets. Efficiency is maintained by freezing most of the LLM and updating only a small fraction of parameters, making the approach practical for deployment with scalable hardware. Overall, the paper advances RF fingerprinting by delivering robust generalization, rapid adaptation with minimal labeled data, and built-in interpretability grounded in local signal structure.

Abstract

Deep neural networks (DNNs) have achieved remarkable success in radio frequency (RF) fingerprinting for wireless device authentication. However, their practical deployment faces two major limitations: domain shift, where models trained in one environment struggle to generalize to others, and the black-box nature of DNNs, which limits interpretability. To address these issues, we propose a novel framework that integrates a group of variable-length two-dimensional (2D) shapelets with a pre-trained large language model (LLM) to achieve efficient, interpretable, and generalizable RF fingerprinting. The 2D shapelets explicitly capture diverse local temporal patterns across the in-phase and quadrature (I/Q) components, providing compact and interpretable representations. Complementarily, the pre-trained LLM captures more long-range dependencies and global contextual information, enabling strong generalization with minimal training overhead. Moreover, our framework also supports prototype generation for few-shot inference, enhancing cross-domain performance without additional retraining. To evaluate the effectiveness of our proposed method, we conduct extensive experiments on six datasets across various protocols and domains. The results show that our method achieves superior standard and few-shot performance across both source and unseen domains.

Figures (7)

• Figure 1: Overview of the proposed RF fingerprinting system. Few-shot inference is enabled when target domain data is available.
• Figure 2: The t-SNE visualization of features under different LLM adaptation strategies. S-$1$ to S-$3$ represent three devices in the source domain, while T-$1$ to T-$3$ denote the same devices in the target domain.
• Figure 3: LoRa and BLE dataset devices.
• Figure 4: Few-shot performance in cross-dataset evaluation. The shaded regions represent the accuracy ranges across multiple unseen domains. O, W, C, L, and B refer to ORACLE, WiSig, CORES, LoRa, and BLE, respectively; O$\rightarrow$W indicates training on ORACLE and testing on WiSig.
• Figure 5: Visualization of learned shapelets and the matched subsequences. S$\#$ denotes the shapelet index, and $t$ indicates the starting time index. The I and Q components are shown in the first and second rows. Blue: real subsequence; red dashed: matched shapelet.