SatForensics: a transformer-based multi-modal framework for cyber-attack detection, attribution, and chain-of-custody forensics in low-earth-orbit satellite communication networks
Year: 2026
Type: article
Abstract: Low-Earth-Orbit (LEO) satellite networks lack public records of labelled cyber-attacks, which makes it hard to evaluate detection, attribution and forensic tools for them. We present SatForensics, a gated multi-modal Transformer that scores each satellite observation for anomalies, attributes injected RF attacks to one of five archetypes (broadband, continuous-wave, swept and pulsed jamming, and Doppler-following spoofing), and records every event in a SHA-256 hash-chained ledger together with a calibrated posterior, a confidence margin and an ambiguity flag. The model fuses RF features of the recorded waterfall, orbital geometry propagated from the observation’s own orbital elements, and link context. We evaluate it on 5,408 real SatNOGS observations from 573 satellites and 379 stations, using satellite-grouped splits, attacks injected after splitting, and 5 runs reported as mean ± standard deviation. SatForensics reaches a detection F1 of 81.0 \(\pm \) 1.4% and an attribution macro-F1 of 96.1 \(\pm \) 0.7%. It outperforms SVM, CNN and LSTM baselines in detection but not Random Forest (83.5 \(\pm \) 1.4%) or an RF-only Transformer (84.3 \(\pm \) 1.6%), and ablations show no reliable contribution of gating, geometry, context or mutual-information alignment on this data. Natural labels are SatNOGS vetting outcomes, not attacks, and attack labels are synthetic. Ledger appends take a median of 10.0 \(\mu \) s and inference takes 0.47 ms on one CPU thread. The code and dataset construction scripts are released.
Source: Scientific Reports
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7213889734
