Lightweight RF Fingerprint Identification using Cross Modal Knowledge Distillation: Learning from Wired to Wireless

Radio frequency fingerprint identification (RFFI) is considered as a feasible zero-trust IoT identification, however, the presence of channel effects in wireless propagation significantly impacts its identification performance. In this paper, we propose a cross modal knowledge distillation based RFFI network (DrffNet) to mitigate the channel effects. Specifically, in our proposed DrffNet network, wired and wireless long range (LoRa) RF signals are jointly used as inputs to the teacher network and student network, respectively. By transferring the RFF knowledge from the hidden layers of the teacher network, the student network can achieve effective RFFI that is resilient to the channel effects. The student network used for RFFI of a similar performance has a parameter size of only 10.68 MB, achieving a 27% reduction compared with the teacher network. We conduct experiments with real-world datasets of the LoRa devices and analyze the feasibility of the proposed network via the ablation experiments. The results demonstrate that our proposed network achieves an identification accuracy of above 95% under various channel environments. Additionally, we show that the proposed network outperforms the benchmark convolutional neural network (CNN) networks on RFFI.