PairWatch perfectly separates real adversarial rendezvous from operational satellite behavior
The claim
PairWatch can detect all documented RPO events while rejecting operational negatives with 100% success across bootstrap resamples
Published as: “PairWatch: Pair-Aware Anomaly Transformer for Detecting Adversarial Rendezvous in Public Satellite Catalogs”
So what
Enables catalog-scale automated screening for adversarial rendezvous with zero false alarms at full recall, fitting within the TLE update cadence.
The five-minute version
PairWatch uniquely achieves a single-threshold separation of real adversarial rendezvous from operational negatives, attaining operational AUROC 1.000 and a positive separation gap of +0.010 that holds in 100% of bootstrap resamples, whereas baselines succeed in at most 19.5% of resamples. The evidence comes from an open benchmark of 359 TLE windows (11 real RPO events) with statistical testing, but performance drops to chance on a helix-formation test and the evaluation covers only a single cross-actor case.
What they found 2
- The property holds in 100% of bootstrap resamples, far exceeding baselines
“PairWatch’s positive gap holds in 100% of resamples, against 19.5% for TranAD, 2.3% for Vanilla AT, and 0% for LSTMAD”
- PairWatch reaches perfect AUROC on all realistic operational contrasts
“PairWatch reaches a perfect 1.00 on every realistic operational contrast-real RPO against clean co-orbits, routine maneuvers, and Starlink neighbors”
What undercuts it the paper’s own results
- All methods, including PairWatch, drop to near-chance AUROC (~0.60) on the TerraSAR-X / TanDEM-X helix formation, a realistic adversarial-like maneuver
- Cross-actor validation is limited to a single Chinese event, leaving broader generalization untested
- Training only uses Clohessy-Wiltshire inspection patterns and not one-shot flybys, and relies on a single simulator
Stated limits 3
- against the TerraSAR-X / TanDEM-X helix formation (∼2 km median range, 0.12 km closest approach), every method we tested degrades to near chance (AUROC ∼0.60). The reason is structural rather than a modeling gap: the helix maintains a quasi-periodic relative orbit where range rate crosses zero twice per orbit with sign-symmetric closing and receding phases-precisely the signature our prior amplifies and the same a deliberate inspection produces. At TLE resolution, the only separating feature would be mission intent, absent from the kinematic stream.
- the cross-actor probe covers a single Chinese event; benchmark-scale evaluation across additional Chinese (SJ-12 / SJ-06F) and commercial (MEV-1 / Intelsat-901) RPO events remains to be conducted.
- the Clohessy-Wiltshire trajectories in training cover sustained co-orbiting inspection patterns but not one-shot flybys with sharp closing-then-receding profiles, and the single-simulator reliance calls for cross-simulator evaluation as new events enter the public record.
What’s new here 4
- Models satellite pairs instead of single objects with a sign-symmetric geometric similarity prior
- Introduces the first public benchmark of real and synthetic RPO windows from Space-Track TLE data
- Evaluates a deployment-relevant single-threshold separability metric rather than aggregate AUROC alone
- Shows that prior neural baselines exploit a synthetic-vs-formation shortcut that PairWatch avoids
Artifacts & methods 12
- PairWatchsystem
- Open RPO benchmark (359 windows)dataset
- Clohessy-Wiltshire synthetic adversarial windowsdataset
- TerraSAR-X / TanDEM-X helix formation setdataset
- LSTM-ADsystem
- TranADsystem
- Vanilla AnomalyTransformersystem
- Isolation Forest (single-satellite mean-element residuals)system
- Range-rate swing scoresystem
- Clohessy-Wiltshire template matchingsystem
- DTW 1-NNsystem
- One-class SVM with engineered pair descriptorssystem