RF Fingerprint Recognition Method Based on Lightweight Deep Learning for Doppler Resilience

Radio Frequency (RF) Fingerprint Recognition has emerged as an area of significant promise within the research community. However, its effective deployment in scenarios characterized by vehicular movement or pedestrian traffic remains a challenge, primarily due to the Doppler effect inducing frequency shifts in RF signals. To counter this, a novel Channel-Independent Spectrum Construction methodology has been proposed, aimed at mitigating the influence of Doppler frequency shifts. This study integrates this innovative approach with advanced data augmentation techniques, thereby pioneering an enhanced methodology for bolstering the resistance of RF signal fingerprint features against Doppler interference. Employing sophisticated deep learning algorithms, this research effectively extracts complex, nonlinear characteristics, thereby capturing the nuances of Doppler variations present within signals. Furthermore, this paper conducts a comprehensive comparison between traditional Convolutional Neural Networks, Deep Residual Networks, and our optimized lightweight PruMobileNet model, an adaptation of MobileNetV1. The empirical results from our study underscore the robustness of our approach in handling Doppler channels, while simultaneously achieving a paradigm shift towards low-cost, environmentally sustainable communication through model lightweighting. Notably, this methodology substantially boosts inference speed and computational efficiency, establishing a new benchmark for reliability in RF fingerprint recognition, particularly in environments with low-speed Doppler effects.