Wavelet-Guided Frequency Decoupling for Channel-Robust UAV RFFI

Abstract

Uncrewed aerial vehicle (UAV) radio frequency fingerprint identification (RFFI) is a promising technique for noncooperative UAV management. However, models trained under line-of-sight (LOS) conditions suffer from significant performance degradation when deployed in complex non-line-of-sight (NLOS) scenarios due to channel dependence. A signal-level analysis suggests that NLOS multipath fading can introduce rapid local variations across the logarithmic spectrogram, which motivates feature-space wavelet decoupling. Accordingly, we propose a wavelet-guided frequency decoupling (WGFD) module that applies multilevel two-dimensional discrete wavelet transform (2D-DWT) to intermediate feature maps, emphasizes approximation responses, and selectively retains complementary detail responses. A multi-stream fusion module (MSFM) integrates identity, local spatial, and wavelet-domain features, while a cross-stage partial aggregation (CSPA) backbone progressively aggregates multilevel representations. Experiments on the DroneRFb dataset show that the proposed method achieves 95.27% accuracy under the LOS $\rightarrow $ NLOS protocol. Subband ablation and feature-space consistency analyses further support the low-frequency-dominant, high-frequency-aware design. The source code is available at: https://github.com/Edith-xx/Channel-Robust-UAV-RFFI