IriSig-Spoof: A Real-World Benchmark for Time-Robust Satellite RF Fingerprinting and Spoofing Detection

Abstract

Low Earth orbit (LEO) satellite Internet is becoming critical communications infrastructure, yet its open wireless links remain vulnerable to satellite impersonation and signal spoofing. Radio frequency fingerprinting (RFF) offers a potential defense by exploiting transmitter-specific hardware imperfections manifested in received signals. However, the reliability of existing satellite RFF methods remains difficult to assess because no unified dataset and benchmark support temporal, open-set, and cross-scenario evaluation. To address this gap, we introduce IriSig-Spoof, a real-world Iridium dataset comprising 5.17 million messages collected from 66 satellites over 32 days, together with software-defined radio (SDR)-generated spoofing signals from indoor and outdoor settings. We further establish three benchmark tasks: temporal robustness evaluation, open-set RFF identification with unknown-signal rejection, and cross-scenario spoofing detection. Experiments using a multi-scale attention convolutional neural network (MACNN) show that temporal robustness varies across configurations, with the best configuration achieving 97.75% average cross-day accuracy. In open-set evaluation, MACNN achieves an area under the receiver operating characteristic curve (AUROC) of 0.9715, while showing that effective unknown-signal rejection does not necessarily ensure reliable identity assignment. Cross-scenario experiments reveal differences at low false-positive rates. IriSig-Spoof provides a reproducible basis for evaluating robust RFF methods under temporal variation and changing attack conditions.