LLRF: Towards Long-Term LoRa Radio Frequency Fingerprint Identification Based on Transfer Learning

Radio Frequency Fingerprint (RFF) identification is an emerging security solution that is expected to guard the flourishing Long Range (LoRa) devices. Noting LoRa's deployment duration ranges from months to years, a fundamental issue that hinders RFF identification for practical LoRa deployment is the long-term performance. In this paper, we propose LLRF, a new approach to improve the long-term RFF identification performance based on transfer learning. We first tailor a series of data processing for LoRa RFF. We then propose an integration of transfer learning and iterative training methods for transfer capability enhancement; it employs experimental data from consecutive days for iterative training to adapt the RFF variations over time. On a real testbed, we evaluate the proposed LLRF across a half year. Experimental results demonstrate that the pre-trained LLRF can achieve an average identification accuracy of 89% after 10 days, 93% after 1 month, and 86% after a half year. Comparison studies further demonstrate the superiority of the proposed LLRF over existing methods and its performances under parameter, environmental, and network architecture variations.