DSC-CAN: a cross-attention network with dual-space calibration for RF-based open-set UAV identification

Abstract The rapid expansion of unmanned aerial vehicle (UAV) applications increases the demand for reliable low-altitude airspace monitoring. Radio-frequency (RF)-based UAV identification provides a passive and non-visual solution by exploiting transmitter-dependent signal characteristics. In open-world deployments, an RF monitoring system needs to recognize known UAV emitters while rejecting previously unseen RF signals. However, open-world RF monitoring remains challenging because known UAV signals often exhibit substantial intra-class variations, while unseen signals can lie close to known-class distributions. These factors make it difficult to preserve known-class discrimination and reliably reject unknown signals under resource-constrained deployments. To address these limitations, we propose the Dual-Space Calibrated Cross-Attention Network (DSC-CAN), a parameter-efficient framework for open-set RF UAV identification. We first employ a compact local–global cross-attention encoder to coordinate fine-grained RF patterns and long-range contextual dependencies, thereby learning compact and discriminative representations. We then introduce class-adaptive prototype-radius calibration to construct class-specific acceptance regions from known training and calibration data, enabling adaptive boundary estimation without using unknown test samples. Finally, we develop dual-space geometric fusion to combine Euclidean prototype-distance evidence with contrastive cosine evidence, thereby suppressing unreliable acceptance decisions when the two spaces provide inconsistent support for a known class. Experiments on multiple known–unknown class partitions show that DSC-CAN achieves strong and stable open-set performance with only 0.187 M parameters. In the primary scenario, DSC-CAN obtains an H -score of 0.9542, an AUROC of 0.9847, an FPR@95 of 0.0368, and an OSCR of 0.9586. Ablation studies further validate the effectiveness of the overall design and the complementary roles of its components.