RF-TCNet: A Lightweight Topology Compression Network for Drone RF Fingerprint Identification

As drones become increasingly prevalent in both civilian and military applications, identifying their Radio Frequency (RF) characteristics is critical for airspace security and drone management. This paper proposes a lightweight network architecture RF-TCNet for drone RF fingerprint identification. A preprocessing method called Energy-Calibration Spectrum Generation (ECSG) is developed, which uses the global maximum amplitude to calibrate the spectrum energy and enhances feature contrast using decibel (dB) scaling transform to generate high-quality training data. Subsequently, the RF-TCNet is used for classification, which has approximately 0.1 M trainable parameters. Its core modules include Dynamic Frequency Attention (DFA) that emphasizes critical frequency elements and Energy Topology Pooling (ETP) that amplifies high-energy regions by eliminating redundant data. Experiments conducted on the DroneRFa and DroneRF datasets show that ECSG improved classification accuracy by 6.14% and 9.85%, respectively, compared to traditional preprocessing methods. With RF-TCNet, we achieve classification accuracies of 99.97% and 94.89% on these datasets while maintaining an extremely low number of parameters. The work improves the performance of drone RF signal recognition through efficient lightweight design and targeted preprocessing methods, providing a potential solution for resource constrained scenarios.