qsh-satellite-opssat-ckks-student

CKKS-linear OPS-SAT segment student

logistic regression over the quadratic expansion (209 terms) of the 18 published per-segment features + sampling

Part of QSMPC-QKD-QHE-AI-Hybrid, a quantum-safe orchestration demo. This is the encrypted-tier model for the satellite_comms use case: a distilled student small enough to evaluate under CKKS. Whether the public demo encrypts THIS model differs by use case; the notes below say what actually runs. It is not the plaintext teacher, and its numbers should be read against that teacher rather than against the published literature. Its agreement with the teacher was measured offline, in plaintext.

Measured performance

Published baselines this is measured against

• Target metric: AUC PR
• Baseline to beat: 0.979 - FCNN, best of the 30 algorithms benchmarked in Ruszczak et al., Scientific Data 12 (2025), on this dataset's own test split (AUC PR 0.979, F1 0.946, MCC 0.932). XGBOD reaches 0.975 and RF+ICCS 0.963.
• Published ceiling: 0.979 - FCNN, the best AUC PR among the 30 baselines Ruszczak et al. benchmarked on this split; no higher published AUC PR was found.

Training data

• Dataset: OPS-SAT-AD
• Licence: CC-BY-4.0
• Source: https://doi.org/10.5281/zenodo.12588359 (licence read 2026-10-31)
• Attribution (CC-BY-4.0): Ruszczak, Kotowski, Evans & Nalepa, 'The OPS-SAT benchmark for detecting anomalies in satellite telemetry', Scientific Data 12, 710 (2025), doi:10.1038/s41597-025-05035-3, CC BY 4.0. Data: Ruszczak, B., Kotowski, K., Nalepa, J. & Evans, D., 'OPSSAT-AD - anomaly detection dataset for satellite telemetry', Zenodo (2024), doi:10.5281/zenodo.12588359 (version 1).

2,123 annotated telemetry segments across 9 channels from ESA's OPS-SAT CubeSat, 20.4% anomalous, with 18 published per-segment features plus a sampling metadata column, and the benchmark's own train/test split. That split is stratified random, not temporal, so scores on it do not measure generalisation over time. Ruszczak et al., Scientific Data 12 (2025). CC BY 4.0, which permits redistributing the data with credit; the per-segment feature table is about 507 KB, the raw segment file much larger. A different task from ESA-ADB: segment classification, not forecasting, so its score is not comparable with the Telemanom model's and neither is the other's baseline. Pinned to version 1 (Zenodo record 12588359), the files this project downloads; version 2 (record 15108715) changes the files and adds a LICENSE file.

Notes and limitations

19 columns is small enough to evaluate under CKKS as a single packed dot product, so the encrypted tier here runs on the published feature set rather than on a reduction of it - and it had room to spare. The student now fits over the QUADRATIC expansion of those features, 209 terms, still one packed dot product at unchanged multiplicative depth, because the expansion is a function of the plaintext input and is computed before encryption. The 209 terms pad to 256 slots, 16 times fewer than the ring's 4,096, so each ciphertext packs 16 rows and the slots set the batching (34 ciphertexts for 529 segments), not a limit on the feature map; the cipher was never the binding constraint here - the feature map was. Measured: minority-class F1 0.8649 -> 0.9043, AUC-PR 0.9275 -> 0.9681 (0.968055 against the teacher's 0.968107, equal at four decimals). The satellite page evaluates this student under CKKS in the browser on all 529 test segments, as a check of agreement with plaintext. It does not decide the page's verdict, and the client computes the expansion and the z-scoring before encrypting.

Honest scope

This model is published as part of a research proof of concept, not as a production system. Numbers above are what this repository measured on the split described, with the code in scripts/train/. Where a figure is carried from the literature rather than measured here, it is labelled as such.