FiCo: A Fingerprinting-Based Two-Step Learning-to-Learn Approach Combing Vibration and 5G Communication for UAV Classification

Unmanned Aerial Vehicles (UAVs) are widely utilized across industries, necessitating robust security measures. Radio Frequency (RF) fingerprinting based UAV identification can enhance UAV authentication and authorization by identifying unique RF signals, bypassing vulnerabilities of software-based methods. However, existing methods often rely on single-domain signal sources, limiting their robustness by environmental factors such as distance, interference, and multi-path propagation. To address the challenges of UAV identification, we propose a novel lightweight two-step learning-to-learn classification approach, FiCo, which integrates multiple data sources from diverse domains, including mechanical vibration and 5 G communications. In the first step, we employ two Extreme Gradient Boosting (XGBoost) models to separately analyze communication and vibration data from UAVs. In the second step, a Logistic Regression meta-network is utilized to jointly learn from the predictions of these two XGBoost models. Experimental results show that the FiCo method boosts the AUC to 0.9792 and the accuracy to 92.59%. This represents a 2% accuracy increase over the Data Combined method and it improves accuracy by 9.3% with communication data alone and by 5.6% with vibration data alone, raising the AUC by 0.088 and 0.029, respectively. This approach reduces computational complexity and requires fewer training samples, enabling faster and more agile UAV identification in practice.