Muhammaduazir69/aeroscope-adsb-anomaly-benchmark

AeroScope ADS-B Anomaly Benchmark v1

A free, openly-licensed benchmark for evaluating ADS-B anomaly and spoofing detectors. It pairs real airborne ADS-B traffic with synthetically injected attacks following the standard taxonomy in the ADS-B security literature, so detectors can be compared on a shared, labelled, reproducible dataset.

• 918 rows — 459 real / 459 injected (balanced)
• 38 documented columns — raw ADS-B fields, integrity fields (NIC/NACp/NACv/SIL), and derived self-consistency features
• Baseline included — every row carries the score from a deployed scikit-learn IsolationForest (ROC-AUC 0.8943; 54.24% detection at a 2.54% false-positive rate, threshold calibrated on held-out real traffic)
• License: CC-BY 4.0
• Project: aeroscope.live · dataset card: aeroscope.live/research · attack reference: aeroscope.live/ads-b-security

How it was built

Real airborne ADS-B traffic was captured live from crowdsourced 1090 MHz receiver feeds across multiple regions and de-duplicated by ICAO 24-bit address. An equal number of attack records were then synthetically injected following the standard taxonomy (cf. Habler & Shabtai, Computers & Security, 2018):

attack_type	rows	what it perturbs
altitude_tamper	115	barometric/geometric altitude (clamped to the Mode-S floor so rows stay physically transmittable)
velocity_tamper	115	ground speed / Mach / airspeed consistency
ghost_kinematics	115	fabricated kinematics for non-existent (ghost) aircraft
integrity_downgrade	114	NIC/NACp/SIL integrity-field downgrade
(none — real traffic)	459	—

Usage

from datasets import load_dataset

# replace Muhammaduazir69 with your Hugging Face username
ds = load_dataset("Muhammaduazir69/aeroscope-adsb-anomaly-benchmark")
df = ds["train"].to_pandas()

X = df.drop(columns=["record_id", "label", "attack_type", "iforest_score", "iforest_is_anomaly"])
y = df["label"]          # 0 = real, 1 = injected attack
print(df["attack_type"].value_counts(dropna=False))

A JSONL version (aeroscope-adsb-anomaly-benchmark-v1.jsonl, one object per row, same fields) is included.

Columns (38)

record_id, label, attack_type, capture_ts_utc, hex, callsign, registration, aircraft_type, category, source, lat, lon, alt_baro_ft, alt_geom_ft, gs_kts, ias_kts, tas_kts, mach, track_deg, mag_heading_deg, track_rate_dps, roll_deg, baro_rate_fpm, geom_rate_fpm, squawk, nic, nac_p, nac_v, sil, rssi_dbfs, seen_s, seen_pos_s, geom_baro_alt_diff_ft, gs_mach_resid_kts, baro_geom_rate_diff_fpm, track_hdg_diff_deg, iforest_score, iforest_is_anomaly.

Full per-column reference is in DATASET_CARD.md.

Honest limitations

• Synthetic injection is a proxy for real attacks, not the real thing.
• Coverage is biased toward busy terminal areas in the capture regions.
• Single-snapshot kinematics (no trajectories) — a trajectory release is planned.
• Enhanced-surveillance fields (ias_kts/tas_kts/mach/mag_heading_deg/track_rate_dps/roll_deg) are broadcast by only ~25–40% of aircraft, so those columns are intentionally sparse — expected, not data loss.
• nac_v is recorded as-broadcast; some encoders exceed the DO-260B nominal 0–4 range.
• ADS-B is public broadcast data; no anonymisation is applied (hex/callsign/registration as broadcast).

Citation

@misc{uzair2026aeroscopebenchmark,
  author       = {Uzair, Muhammad},
  title        = {AeroScope ADS-B Anomaly Benchmark v1},
  year         = {2026},
  publisher    = {aeroscope.live},
  note         = {CC-BY 4.0},
  url          = {https://aeroscope.live/research}
}

Plain text: Uzair, M. (2026). AeroScope ADS-B Anomaly Benchmark v1. aeroscope.live. CC-BY 4.0. Maintained by Muhammad Uzair, Department of Computer Science, COMSATS University Islamabad (ORCID 0009-0002-4104-2680).