Secure and Intelligent Sensing in Unmanned Aerial Vehicles: A Semi-Supervised Modulation Recognition Framework

Unmanned aerial vehicles (UAVs) leverage various wireless communication systems for efficient operation and data transmission. In these systems, automatic modulation recognition (AMR) plays a crucial role in enhancing signal detection and recognition, ensuring secure and reliable communications. However, the high cost and large amounts of annotated training data required pose significant challenges, particularly in resource-constrained UAV systems. To address these challenges, we propose a semi-supervised AMR (Semi-AMR) method incorporating virtual adversarial training (VAT), which introduces a novel measure of local smoothness in the conditional label distribution based on adversarial perturbations, eliminating the need for pseudolabels. In addition, we introduce an improved proxy-based metric learning (ML) loss to establish a semantic distance metric between radio signals, particularly for unlabeled data. This allows the network to project signals into an embedding space where semantically similar signals are grouped together. The proposed method is evaluated on three real-world datasets: RML2016.10a, RML2016.10b, and RML22. Even with only 25% of the data labeled, the method achieves an identification accuracy of 83.09% at 0-dB SNR in RML2016.10a, demonstrating the robustness and effectiveness of our approach in resource-constrained UAV systems, especially in low signal-to-noise ratio environments.