Cross-Domain Open-Set UAV Identification via Multi-View RF Signal Fusion

Abstract

Radio Frequency (RF)-based Unmanned Aerial Vehicle (UAV) identification systems typically assume that training and test data are collected under matched conditions, and that every test sample belongs to a known UAV type. However, practical deployments often face domain shift and unknown UAV types. In this paper, we study a cross-domain open-set setting where known and unknown RF data are captured under different environments. To address these challenges, we propose a unified pipeline that learns a boundary-aware multi-view RF embedding, detects unknown UAVs via a two-stage cascaded detector, and clusters detected unknowns for discovery. The detector first performs coarse filtering with calibrated confidence scores, then applies prototype-distance and density verification for imbalanced scenarios. Experimental results demonstrate 98.89% validation accuracy on known classes, 90.65% F1-score for unknown detection (90.23% precision, 91.08% recall), and 88.14% clustering purity for unknown discovery.