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

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.

💡 Research Summary

The paper addresses two critical challenges that have limited the deployment of deep‑learning‑based radio‑frequency (RF) fingerprinting: (1) severe performance degradation when a model trained in one environment is applied to another (domain shift) and (2) the black‑box nature of deep neural networks that hinders interpretability. To overcome these issues, the authors propose a novel architecture that synergistically combines variable‑length two‑dimensional (2‑D) shapelets with a pre‑trained large language model (LLM).

Shapelets are learnable subsequences that operate on both the in‑phase (I) and quadrature (Q) components simultaneously, capturing fine‑grained local temporal patterns that arise from hardware imperfections unique to each device. By learning a compact set of discriminative shapelets, the system provides intrinsic, human‑readable explanations of why a particular device is identified, something that post‑hoc methods cannot guarantee.

The LLM (e.g., a BERT‑style encoder) is repurposed as a powerful feature extractor for RF data. Raw I/Q samples are first tokenized and projected into the LLM’s input space via a lightweight embedding layer. Most of the LLM’s parameters remain frozen; only positional embeddings and layer‑norm weights are fine‑tuned, dramatically reducing training cost while preserving the model’s extensive pre‑training knowledge. This enables the LLM to capture long‑range dependencies and global contextual information that shapelets alone cannot represent, thereby improving cross‑domain robustness.

For few‑shot inference, the authors adopt a prototypical network paradigm. A prototype vector for each class (device) is computed in the LLM’s embedding space using a small number of labeled examples. During inference, the combined representation (shapelet similarity + LLM embedding) is compared to these prototypes, allowing the system to adapt to new devices or environments with as few as five samples and without any additional retraining.

Extensive experiments were conducted on six publicly available datasets covering Wi‑Fi, LoRa, and Bluetooth Low Energy (BLE) protocols, each with multiple source and target domains. The proposed method consistently outperformed state‑of‑the‑art deep learning baselines. In the source domain, accuracy improvements ranged from 4 % to 6 % over the best existing models. More strikingly, when evaluated on unseen target domains without any domain‑adaptation step, the method achieved an average gain of over 10 % in accuracy, demonstrating strong generalization. In few‑shot scenarios, using fewer than five labeled samples per new class, the system retained >90 % of its full‑data performance, highlighting its practical utility for real‑world deployments where labeling is costly.

Interpretability was validated by visualizing the learned shapelets, which revealed distinct I/Q patterns corresponding to specific hardware anomalies of each device. These visual explanations align with engineering intuition, offering confidence to security analysts and system designers.

In summary, the paper makes three major contributions: (1) it is the first work to integrate a pre‑trained LLM into RF fingerprinting, leveraging the LLM’s transfer learning capabilities to mitigate domain shift; (2) it introduces a shapelet‑enhanced fine‑tuning framework that provides built‑in, model‑intrinsic interpretability without expensive retraining; and (3) it demonstrates a prototype‑based few‑shot inference mechanism that enables rapid adaptation to new devices or environments. The combination of local shapelet explanations, global LLM representations, and efficient few‑shot learning constitutes a significant step toward deployable, trustworthy RF authentication systems.