OrbID: Identifying Orbcomm Satellite RF Fingerprints
Cédric Solenthaler, Joshua Smailes, Martin Strohmeier
TL;DR
This work addresses securing Orbcomm satellite communications through RF fingerprinting learned from IQ data. It introduces a CNN-based embedding trained with triplet loss on a large, multi-location Orbcomm dataset of 8,992,474 packets, including a payload-anonymizing preprocessing step and data augmentation to promote cross-SDR transferability. The approach achieves a high spoofing-detection performance (ROC AUC up to 0.98) and demonstrates discrimination within the constellation (ROC AUC 0.53) and robust cross-SDR/generalization properties, with ~30 ms per fingerprint inference. The results indicate that ground-segment physical-layer authentication can significantly enhance security for legacy satellite systems, with practical implications for deployment and future multi-constellation extensions.
Abstract
An increase in availability of Software Defined Radios (SDRs) has caused a dramatic shift in the threat landscape of legacy satellite systems, opening them up to easy spoofing attacks by low-budget adversaries. Physical-layer authentication methods can help improve the security of these systems by providing additional validation without modifying the space segment. This paper extends previous research on Radio Frequency Fingerprinting (RFF) of satellite communication to the Orbcomm satellite formation. The GPS and Iridium constellations are already well covered in prior research, but the feasibility of transferring techniques to other formations has not yet been examined, and raises previously undiscussed challenges. In this paper, we collect a novel dataset containing 8992474 packets from the Orbcom satellite constellation using different SDRs and locations. We use this dataset to train RFF systems based on convolutional neural networks. We achieve an ROC AUC score of 0.53 when distinguishing different satellites within the constellation, and 0.98 when distinguishing legitimate satellites from SDRs in a spoofing scenario. We also demonstrate the possibility of mixing datasets using different SDRs in different physical locations.