Radio Frequency-Based Unmanned Aerial Vehicles Identification with Multi-Domain Self-Attention

The revolutionary development of the unmanned aerial vehicles (UAVs) has brought a paradigm shift towards the high-tech era for modern people's lifestyles. However, current UAV regulatory techniques have notable deficiencies in terms of recognition performance and resource consumption. In con-junction with the abuse of UAVs, these issues have raised critical concerns about personal privacy and public security. This paper proposes a novel methodology for effectively identifying UAVs via radio frequency (RF) signal. This methodology is herein referred to as the Identification Model with Multi-Domain Self-Attention (IM-MDSA). The short-time Fourier transform (STFT) is firstly utilized to generate the time-frequency spectra (TFS), thereby decoupling the UAV signals from the same-band interference and ambient noise. Secondly, two feature extractors based on improved self-attention mechanisms are developed to capture the variability in RF fingerprints across the space, time, and frequency domains. Finally, a lightweight classifier is desig-ned based on an improved convolutional neural network (CNN). Numerous experiments conducted on DroneRFb-Spectra dataset show that IM-MDSA achieves an accuracy rate of 99.64% for 12 different categories of signal while concurrently requiring as few as 0.52 million parameters. This results underscore the efficacy of IM-MDSA compared to alternative studies, thereby providing a promising method for the future UAV identification.