LRF-KDTL: Lightweight RF Fingerprint Identification Based on Knowledge Distillation and Transfer Learning

Physical-layer authentication (PLA) provides an important approach to achieve secure access by exploiting the inherent hardware characteristics of devices. Radio frequency (RF) fingerprinting-based PLA emerges as a cost-effective solution to counter identity spoofing threats. Although deep learning (DL) has been widely adopted for RF fingerprint extraction and device identification, existing methods typically require complex DL models with massive parameters to achieve high accuracy, making applying them to resource-constrained devices impractical. Furthermore, most approaches focus on closed-set recognition scenarios, failing to address practical environments containing unknown devices. To overcome these limitations, we propose LRF-KDTL, a novel lightweight DL-based RF fingerprint identification method leveraging knowledge distillation and transfer learning. Specifically, LRF-KDTL employs a teacher-student framework for feature distillation and subsequently retrains the student model through transfer learning, achieving precise device identification with significantly reduced complexity. Extensive experiments in open-set scenarios demonstrate that the proposed method reaches 90.21% identification accuracy while reducing model parameters by 92.36%, outperforming existing approaches by 3.13%. These results validate the effectiveness and superiority of LRF-KDTL for practical deployment.