MAML-Resnet for a Low-Data Approach to Unmanned Aerial Vehicle Identification
Mingdian Li, Anze Shen, Joshua Le-Wei Li, Yang Li · 2025
Unmanned aerial vehicles (UAVs) have revolutionized various industries, but their unregulated use poses significant security risks, necessitating reliable identification methods. Radio frequency (RF) fingerprinting offers a promising solution by leveraging hardware-specific signal imperfections, yet it faces challenges such as data scarcity and poor adaptation to new classes in low-data scenarios. This paper proposes a MAML-ResNet framework that combines Model-Agnostic Meta-Learning (MAML) with the ResNet-18 architecture to address these challenges. The framework converts raw RF signals into time-frequency diagrams via short-time Fourier transform (STFT) for hierarchical feature extraction by ResNet, while MAML enables rapid adaptation to new UAV classes with few samples through bi-level optimization. Experiments on the DronerFa dataset demonstrate that the proposed method achieves$\mathbf{7 6. 4 9 \%}$accuracy in 5-way 5-shot tasks, outperforming traditional ResNet (59.74 %) and shallow CNN baselines by over 15 %. Future work may explore multi-domain feature fusion, open-set recognition, and lightweight architectures for real-world deployment.