One-step Framework for RF-Based Drone Identification

With the popularity of drones, incidents that threaten public safety caused by drones occur more frequently. In drone control and countermeasures scenarios, it is needed to further identify drones of the same type based on type recognition in order to determine the specific source of the signal. Existing schemes are multistep, which increases the complexity and overhead of post-maintenance. In this work, we propose a joint scheme based on RE fingerprinting where the type and individual belonging to the same type are identified in one step. Specifically, we adopt short-time Fourier transform to calculate the spectrogram of the received signal and label the types and individuals of the drone signals on the spectrogram to develop a data set. Based on the data set, we train a YOLOv5 neural network to extract detailed RF characteristics to identify which type and individual the signal belongs to. To further improve the identification performance, we introduce grayscale image colorization. Furthermore, we use median filtering to reduce noise and protect image edge information. Results on data set composed of real signals from 18 drones of 9 types show the classification accuracy of our proposed framework reaches over $98\%$ at saturation and $3\%$ higher than the method without grayscale image colorization.