Large Language Model Enabled Lightweight RFFI for 6G Edge Intelligence

The radio frequency fingerprint identification (RFFI) is promising which exploits the hardware defects inherent to realize the zero-trust Internet of Things (IoT) security. Considering the scalability, the data imbalance and the training overhead for the current deep learning (DL) based RFFI, in this paper, we combine the large language model (LLM) and propose a BERT-LightRFFI framework, to enhance the zero-trust edge IoT security. Specifically, we pre-train a BERT model with the unlabeled data via self-supervised learning and obtain a powerful RFF feature extractor. Then, we use the knowledge distillation to inherit the BERT learn-gene to the lightweight BERT-Light model, and then fine-tune the BERT-Light model and a classifier with a few-shot labeled wireless data. In the experiments, we use a large-scale real-world LoRa dataset to evaluate the performances of the proposed framework and propose the interesting insights. The results prove the effectiveness of the proposed framework, achieving an accuracy of 97.52% with less model parameters and computation amount in the presence of multipath fading and Doppler effect, which is better than the benchmark methods.