Radio Frequency Fingerprint Identification Based on Variational Autoencoder for GNSS

Interference against global navigation satellite system (GNSS) is threatening its reliability. Radio frequency fingerprint identification (RFFI) emerges as a physical-layer security solution that can effectively identify genuine transmitters. However, external noise in the transmission is not conducive to maintaining the robustness of the RFFI, and the deep neural networks (DNNs) or large datasets for boosting robustness will consume excessive resources. To this end, this letter proposes a lightweight RFFI scheme based on variational autoencoder (VAE) and long short-term memory (LSTM) for real-field GPS signals collected in Nuremberg. The VAE aims to denoise and reconstruct the RFFs, thereby improving the identification accuracy and reducing feature dimensionality. LSTMs can extract the RFF features without any pretransformation and avoid the problem of gradient vanishing or gradient exploding. Numerical results demonstrate that our model can yield an identification accuracy of up to 95.68% on postcorrelation GPS data at low complexity.