TrainTrack Dataset

The TrainTrack Dataset is a collection of smartphone inertial sensor recordings captured by passengers during real train journeys. It was collected for the paper "On the Same Track: Privacy Leaks in Electric Rail Transport via Magnetic Fields" (CCS '26), which studies how the magnetic field of electric trains leaks through the smartphone magnetometer.

What the dataset contains

The dataset consists of two parts:

traintrack-dataset/ — 487.5 hours of labeled train travel. Recordings from participants who regularly commute by train, covering 217 journeys, 342 trips, and 2,001 station-to-station segments across 246 unique stations. The data is organized hierarchically: a journey is one continuous recording session of a participant's travel, a trip is a single train ride between two changes within a journey, and a segment is one station-to-station leg of a trip. 357 segments were recorded by two or more devices simultaneously aboard the same train, making them usable for colocation studies.

Every segment is annotated with the train type and whether the train runs on a 16.7 Hz AC overhead line. The distribution by train type, given as total hours and segments with the 16.7 Hz electric share in parentheses:

• Long distance — 420.1 h, 862 segments (thereof 16.7 Hz electric: 408.0 h, 840 segments)
• Regional — 43.5 h, 346 segments (thereof 16.7 Hz electric: 36.4 h, 281 segments)
• Urban — 19.7 h, 586 segments (thereof 16.7 Hz electric: 3.7 h, 88 segments)
• Metro — 4.2 h, 201 segments (thereof 16.7 Hz electric: none)
• Tram — 0.1 h, 6 segments (thereof 16.7 Hz electric: none)
• Total — 487.5 h, 2,001 segments (thereof 16.7 Hz electric: 448.1 h, 1,209 segments)

no-trainride-dataset/ — everyday activity outside of trains. Recordings collected by the same participants while carrying their smartphones through ordinary daily life, serving as negative examples for transport mode detection. These traces are annotated with activity labels from Google's Activity Recognition API and cover roughly 78 hours in total, distributed over stationary periods (~108,000 s), walking (~33,000 s), cycling (~22,000 s), automotive travel (~30,000 s), and unclassified intervals (~90,000 s).

Sensors

Recordings were made with the Sensor Logger app. Each recording is a set of CSV files, one per sensor, sharing a common journey-relative time base:

• Magnetometer (calibrated and uncalibrated)
• Accelerometer (calibrated, uncalibrated, and total acceleration)
• Gyroscope (calibrated and uncalibrated)
• Gravity, Orientation, Game Orientation, Geomagnetic Orientation, Compass
• Barometer
• Per-recording device metadata (model, platform, app version, sensor list, sample rates)

Code and documentation

Detailed instructions for downloading, extracting, and working with the dataset, along with the full implementation and experiments from the paper, are available at:

https://github.com/mlsec-group/magtrack

Citation

@inproceedings{czybik2026magtrack,
  title     = {On the Same Track: Privacy Leaks in Electric Rail Transport via Magnetic Fields},
  author    = {Czybik, Stefan and Hanfeld, Pia and Rieck, Konrad},
  booktitle = {Proceedings of the 33rd ACM Conference on Computer and Communications Security (CCS)},
  year      = {2026}
}