UAV Individual Identification Based on Multi-Modal Fusion of RF Fingerprints

Abstract

The widespread adoption of Unmanned Aerial Vehicles (UAVs) across military and civilian domains has raised urgent concerns about airspace security and UAV traffic management. Among the various technical challenges, identifying specific individual UAVs - distinguishing between different physical drones of the same brand and model - under diverse and challenging signal conditions remains a key unresolved problem. Radio Frequency Fingerprinting (RFF) provides a non-cryptographic means of device identification, but existing approaches predominantly rely on single-modality representations, which offer limited discriminative power and degrade rapidly under non-lineof-sight (NLOS) and line-of-sight (LOS) conditions. This paper introduces a gated dual-modality fusion framework that combines domain-specific expert features with time-frequency signal representations for UAV individual identification. We design a feature set tailored to UAV RF emissions and employ a dualbranch time-frequency encoder with Squeeze-and-Excitation (SE) attention. A vector-level gated mechanism fuses the two modality streams, adaptively balancing their contributions for robust integration. Evaluated on the DroneRFb-DIR dataset under both LOS and NLOS conditions, our model achieves 96.05% accuracy under LOS conditions and 85.74% under NLOS conditions, consistently surpassing existing fusion strategies (concatenation and attention-based fusion) and single-modality baselines.