An Efficient Dual-Branch Network and Multimodal Fusion Framework for Drone Identification

In recent years, the widespread adoption of drones, while offering convenience, has also led to significant security challenges such as illegal intrusions and privacy violations, creating an urgent need for reliable identification and classification systems. A primary obstacle to achieving this reliability is the high similarity of radio frequency (RF) signals among different drone models, which often leads to misclassification. In this study, we propose the DS-UAVNet, a network that employs a dual-branch architecture to independently process complementary information from the time and frequency domains, thereby preventing information loss. Within this network, a designed parallel convolution module efficiently extracts multi-scale features while significantly reducing model complexity. To address the inherent vulnerabilities of the single-modality drone identification system, we further design M-DS-UAVNet, a multimodal framework that enhances identification robustness by leveraging a transfer learning strategy to fuse audio and RF features. Evaluations show that DS-UAVNet achieves accuracies of 98.74% and 98.56% on the public DroneRF dataset for drone classification and operation mode recognition, respectively, outperforming existing methods. Moreover, the M-DS-UAVNet framework achieves 100.00% and 99.78% accuracy on the constructed multimodal dataset, validating the effectiveness of the multimodal fusion strategy for building identification systems.