Receiver-Agnostic Radio Frequency Fingerprinting Based on Two-stage Unsupervised Domain Adaptation and Fine-tuning

Jiazhong Bao,Xin Xie,Zhaoyi Lu,Jianan Hong,Cunqing Hua

DOI: https://doi.org/10.1109/globecom54140.2023.10437074

2023-01-01

Abstract:Radio frequency fingerprint identification (RFFI) has been widely studied as a physical layer security scheme for device identification and authentication in wireless scenarios, such as Internet of Things (IoTs), industrial wireless networks, Internet of Vehicles (IoV), etc. Typical RFFI approaches train a model at the receiver to extract hardware defects of the transmitter RF front-end using a deep learning-based method and achieve classification. However, few works have taken into account its shortage in multiple-receiver scenarios, where the identification accuracy significantly decreases when migrating a model trained on the known receivers to the new ones, directly. In this paper, we propose a novel cross-receiver RFFI scheme to improve the performance and the generalization of the fingerprinting classification tasks on new receivers. This scheme tackles the shortage by two means: 1) we extract receiver- independent features using global domain adaptation based on adversarial training and relevant subdomain adaptation based on local maximum mean discrepancy (LMMD); 2) The performance is further improved by fine-tuning on few labeled samples when domain adaptation is not effective. The second mechanism brings in significant performance advantage, without a large amount of labeled data on new receivers. Experimental results on public datasets show the outstanding performance of the proposed scheme in cross-receiver scenarios.