Backdoor Attacks Against Low-Earth Orbit Satellite Fingerprinting

With the increasing popularity of Low-Earth Orbit (LEO) satellite communication, its security problems become im-portant. Traditional cryptographic authentication schemes may be outdated and fragile. Radio frequency (RF) fingerprinting emerges as a robust physical layer authentication method that dis-cerns the unique characteristics of each transmitter. Additionally, deep learning-based fingerprinting systems gain more attention as spoofing countermeasures, owing to the formidable capabilities of deep neural networks. However, the inherent vulnerabilities of deep neural networks bring risks to the fingerprinting system. To investigate backdoor attacks on LEO satellite fingerprinting, we first assess the classic poisoning-based backdoor attack. To make the backdoor attack more practical, our study includes two existing common fingerprinting methods, namely supervised learning-based and few-shot learning-based. Furthermore, we also evaluate data-free backdoor attacks, considering that satellite data may be difficult to access and modify by attackers. Our experimental findings reveal that deep learning-based fingerprinting approaches are susceptible to backdoor attacks. In addition, we also demonstrate that these attacks can evade the existing detection approach.