Real-world UAV recognition based on radio frequency fingerprinting with transformer

Many unmanned aerial vehicles (UAVs) require the installation of automatic dependent surveillance-broadcast (ADS-B) transponders to facilitate their daily management. However, since ADS-B transponders do not have a good security mechanism, they introduce problems including impersonation, spoofing, and private changing of the registration number, making UAV surveillance inconvenient. Radio frequency fingerprinting (RFF) recognition is carried out by utilizing the fact that different electronic devices in a given transponder will affect the transmitted signals, resulting in the formation of RFF features that are unique to the transponder and difficult to forge. Therefore, in this work, a deep learning architecture is proposed to classify UAVs based on ADS-B signals, and a multi-head self-attention RFF recognition model is constructed using variational mode decomposition (VMD) of the preamble data and a transformer encoder for validation. The model achieves better results in terms of noise, Doppler shifting, and multipath effect interference. This method demonstrates that the transformer architecture of natural language processing, combined with appropriate data preprocessing methods, can also be used for RFF recognition, and provides advantages in accuracy and robustness (67.83% vs. 64.17%).