AI Generated Wireless Data for Enhanced Satellite Device Fingerprinting

While Low Earth Orbit (LEO) satellite communications have attracted more and more attention recently, its security issues for wireless communications become greatly important. Implementing Radio Frequency (RF) fingerprinting by leveraging In-phase/Quadrature-phase (I/Q) satellite data is a reliable way to enhance wireless communication security. However, this method still faces two challenges. The first is the high cost of acquiring I/Q data from LEO satellites, and the second is the wide variation of data collected in different periods. To solve these two challenges, we develop a modified Variational Autoencoder (VAE) with a tailored loss function to generate LEO satellite data. Then, to achieve a multi-period model generalization, a cosine similarity classifier is incorporated into prototypical networks for facilitating few-shot learning at LEO satellite data in different periods. In our experiments, the proposed model can achieve a 50-classification accuracy of 99.80% in a single period with virtual satellite data generated by the modified VAE, which is higher than the classification result of 98.05% using only real data. Besides, using few-shot learning, our results demonstrate the effectiveness of model adaptation, obtaining a significant improvement in classification accuracy, from 2% to 77% of the original model.