GIST — GNSS Interference Spatio-Temporal network
▶ Live demo & slides: https://mevichitra.github.io/gist-gnss-interference/
Results in the deck and demo are interim (seed 0); the final multi-seed run is pending.
Detect, classify (jamming vs spoofing) and localize GNSS interference from crowdsourced ADS-B,
with no manual labels. Final-year deep learning project + paper. Full design: PLAN.md;
decisions and dead ends: LOG.md.
Pipeline
| Step | Command | Output |
|---|
| Download + parse days (adsb.lol, ODbL) | uv run python -m gist.data.fetch_days [YYYY-MM-DD ...] | data_store/points/<day>/ |
| Features + weak labels | uv run python -m gist.labels.build [days] | data_store/steps/, data_store/windows/ |
| Scene datasets (real + simulated emitters) | uv run python -m gist.train.dataset <days> --real 300 --sim 300 | data_store/scenes/<split>/ |
| Train GIST | uv run python -m gist.train.train --train train --val val --name gist | runs/gist/ |
| Baselines | uv run python -m gist.train.eval_baselines --train train --test test --out runs/baselines.json | JSON metrics |
| Tests | uv run pytest | |
Splits are by day: train ≤ 2025-06, val 2025-07..2025-12, test ≥ 2026-01.
Layout
src/gist/data download (resumable, throttle-aware), trace parser, regions
src/gist/features physics-consistency step features, 10-min windows
src/gist/labels labeling functions, Dawid–Skene label model, build script, event catalogue
src/gist/sim RF propagation (NumPy/PyTorch), jammer/spoofer injection, scene sampling
src/gist/models GIST (Transformer + spatial attention + physics localizer), baselines, NLS
src/gist/train scene tensors, datasets, training, evaluation
paper/ LaTeX (IEEE), bibliography, venues
Data and licences
- adsb.lol globe_history — ODbL-1.0. Derived datasets we release are ODbL; cite "ADSB.lol and contributors".
- ADS-B Exchange samples are not used (evaluation-only terms). See DATA_LICENSE.md.
- OpenSky Network (Trino) — requested 2026-10-27; used only for events before 2023 (e.g. Denver KDEN 2022).
Code: MIT.
