AROS-S (Autonomous Real-time Onboard Security for Satellites) is a lightweight, one-class anomaly detector that runs on the spacecraft payload and inspects every telemetry packet as it is produced, so that abnormal behaviour is caught onboard rather than minutes later on the ground. Because no public dataset of real in-orbit attacks exists, AROS-S does not train a classifier; it learns a model of normal spacecraft state and treats any departure from it as suspicious. The detector has three layers with different blind spots: two partitioned Isolation Forests catch sudden out-of-range points, a 38-parameter autoencoder catches packets that break the learned feature relationships, and a window autoencoder catches slow drift that no single packet reveals. All layers are served as ONNX graphs verified by a SHA-256 integrity check before loading, and a simulated, signed response loop closes the detect-and-recover cycle.
On a combined synthetic and NASA SMAP normal regime, AROS-S raises no false alarms on the baseline, reduces a previously rejected NASA regime's false-alarm rate from 100 percent to about 6 percent, and separates normal from attack reconstruction error by about three orders of magnitude. On the real NASA SMAP benchmark the per-channel autoencoder reaches an F1 of 0.70, matching the strongest classical one-class baseline at the smallest model size and lowest per-window latency of the methods compared. The paper also assesses what a real CubeSat deployment would require.
Preprint submitted to the IEEE International Conference Automatics and Informatics (ICAI 2026), Varna, Bulgaria. Source code and data are publicly available.