Multi-Scale Additive Hybrid Network (MAH-Net) for Robust Open-Set Unmanned Aerial Vehicle Radio-Frequency Recognition

Abstract

As low-altitude commercial activity expands, unmanned aerial vehicles (UAVs) are increasingly used in logistics, inspection, emergency response, and urban services, bringing a growing need for reliable supervision. Passive radio-frequency (RF) sensing is well suited to this task because it detects controller emissions without target cooperation and remains effective under occlusion, adverse weather, and non-line-of-sight conditions. In practice, however, RF recognizers face an open-set problem: they must identify known UAV controllers while rejecting unseen UAV types, protocol variants, and interference signals absent from training. We propose Multi-Scale Additive Hybrid Network (MAH-Net), a lightweight open-set RF recognition framework that maps compact spectrogram embeddings to calibrated known/unknown decisions. MAH-Net combines a multi-scale backbone with Squeeze-and-Excitation Block (SE-Block) and Convolutional Additive Self-Attention Block (CAS Block) modules, ArcFace angular-margin supervision, and class-conditional Mahalanobis-distance rejection with thresholds estimated only from known calibration data. A lightweight hierarchical clustering step then organizes rejected samples for post-rejection analysis. On the public DroneRFb-Spectra benchmark, with 18 UAV classes treated as known and 6 disjoint classes held out as unknown, MAH-Net achieves 93.02% True Known Rate (TKR), 97.77% True Unknown Rate (TUR), 99.96% Known Precision (KP), and 97.14% Unknown Purity (UP) using only 1.25M parameters. The results point to a favorable safety-efficiency trade-off for practical open-set UAV RF monitoring.