Label Budget Trade-offs and Calibration Transfer in GNSS Spoofing Detection
Learning-based global navigation satellite system (GNSS) spoofing detection requires labels both to learn a detection score and to set its alarm threshold. When these information requirements and their acquisition sources are combined, changes in model discrimination can be confused with changes in threshold calibration. We address this evaluation problem through paired experiments that vary calibration size and source while keeping model scores fixed, and compare training--calibration allocations at an equal total label count. The primary SatGrid experiment uses 12,000 correlator observations from three recording cohorts, 135 fitted models and 3,240 threshold conditions. At a budget of 200 labels, a nominal false-positive rate of 5\%, and separate-source normal calibration, increasing training support from 5 to 50 observations per class reduces the metric classifier's false-negative rate from 63.60\% to 20.30\%, while increasing its false-positive rate from 8.83\% to 15.13\%. The allocation therefore exchanges missed detections for false alarms. Applying the unchanged models and thresholds to 4,000 observations from an additional cohort produces false-negative rates of 99.90\% for random forests and 99.95\% for the metric classifier at the intermediate allocation. Reversed score ordering distinguishes this failure from threshold miscalibration alone. These results show that an allocation advantage within the primary cohorts can disappear under source transfer. Evaluation should therefore account for both training and calibration labels and verify score ordering and both conditional error rates on separate acquisition sources.
Authors
- yi zhao
- bo xie
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11234387/v1
- Primary Topic
- GNSS positioning and interference
- Type
- preprint