MetaRF-Net: Robust Few-Shot UAV Identification via Uncertainty-Guided Multi-Gaussian Prototypes

Abstract

Uncrewed aerial vehicles (UAVs) are increasingly deployed in civilian and industrial environments, raising concerns regarding airspace security and reliable UAV monitoring. Radio frequency fingerprint identification (RFFI) provides a device-level solution, but existing deep-learning-based methods typically require abundant labeled data and generalize poorly to newly encountered devices. This letter proposes MetaRF-Net, a few-shot UAV RFFI framework trained via meta-learning to enable rapid adaptation to unseen UAVs. The approach integrates physically consistent radio frequency (RF) augmentation, a hybrid spectral–temporal backbone, and an uncertainty-guided multi-Gaussian prototype network to improve feature quality and class modeling under limited data. Experiments on real UAV RF datasets show that MetaRF-Net consistently outperforms existing few-shot RFFI methods, demonstrating its suitability for practical RF-based UAV monitoring.