RF-Based Drone Detection Under Open Set Setting

As the explosive increase of drone classes, most existing radio frequency (RF) based drone detection methods belong to the supervised learning approaches, which brings the risk of misclassifying unknown classes. To address this problem, an open set signal recognition method for unknown drone detection is proposed to deal with such a scenario where testing classes do not exactly match training classes. In particular, dilated convolution layers with multi-level receptive fields are used in the process of signal semantic construction, which facilitates feature extraction by jointly exploiting the image-transmission signals and the flight control signals of drones. Besides, an outlier analysis-based classifier is designed in the semantic classification process, which relaxes the necessity of manually setting the bounding thresholds. Furthermore, a newly established real-world dataset DroneRF alpha-Spectra is also released, which includes 12 drone classes with a total number of 8334 samples. Experimental results demonstrate that the proposed method outperforms the compared detection methods, achieving the highest detection rate of 99.26% for unknown drones.