Field noteNov 16, 20266 min read1 source

flyover-scrub strips aircraft Doppler so ADS-B cannot geolocate videos

In one minute

The MIT-licensed flyover-scrub command-line tool removes aircraft Doppler signatures from audio and video and includes a reference scorer to test whether ADS-B-based geolocation still works. Its significance is practical: it treats a subtle location leak as something to measure against an attack, while its author is explicit that the defense is not a guarantee.

What changed: A new tool aims to remove aircraft Doppler signatures from audio and video to thwart ADS-B-based geolocation, and includes a scorer to test resistance.

What happened

The flyover-scrub GitHub repository describes the project as a defense against locating recordings by matching aircraft Doppler curves with public ADS-B flight tracks. The attack works even when a plane is inaudible: a long-window spectrogram can reveal the curve, and a scorer can compare it with modeled tracks from candidate locations. The project includes a reference attack scorer and a trace-and-fit variant. It is MIT-licensed and has a command-line interface: flyover-scrub input.mov -o output.mov scrubs audio, copies video, and strips metadata; flyover-scrub --check input.mp4 reports long tonal ridges without writing a file.

For each short-time Fourier transform bin in the protected band, defaulting to 20–2,000 Hz, the tool treats background below 12 dB above the local floor differently from foreground. It resynthesizes background with fresh random-phase noise at the same local power, removes long, slowly gliding narrowband ridges by replacing them with floor-level noise, and passes other foreground—such as speech, music, and transients—untouched. Level corrections use one whole-file number per frequency, rather than time-varying correction that might restore the identifying structure. The tool also reports grid hum, or electrical network frequency, which can reveal recording time and grid; it removes that hum by default, and --keep-hum preserves it.

The README reports two kinds of evaluation. In tests, a physically modeled flyover was mixed into two real iPhone recordings. With an ADS-B-like track, a clock offset of 2.5 seconds, and a 10-by-10-kilometer search box, raw recordings with an inaudible −30 dB flyover produced location errors of 0–461 meters; the true location ranked first in four of six runs. In 12 scrubbed runs, the error was 1.1–5.1 kilometers and the true location ranked between 351 and 2,087, described as no better than chance. A separate account in the repository says that, on the author's own iPhone videos and public ADS-B data, an inaudible Cessna located an indoor selfie to 61–134 meters, and a helicopter heard from a parked car was matched to the right aircraft among 18.

The repository also describes tests against two real ADS-B-matched flyovers: a Cessna and an R44 helicopter left no traceable line after scrubbing. It says the attack survived 32 of 34 codec and filter chains, and that TikTok passed faint tones through untouched. To audit a recording, users can capture current nearby aircraft with flyover-capture and then run flyover-audit; the capture tool makes a small request every five seconds to the free adsb.lol community API, with backoff. Because the API records only current flights, capture must start before recording. The audit needs the true location; if a file's creation time has been rewritten, users must pass --start explicitly. The package does not provide bulk or historical flight downloads.

What the archive adds

New Runtime's earlier coverage documented how a separate tool, doppler-locate, geolocates aircraft audio by matching Doppler curves to ADS-B tracks (November 5, 2026). The next day, an earlier New Runtime note described a video location identified from faint aircraft sound and public flight data. flyover-scrub is a new defensive event on that known privacy thread, with an implementation and evaluation aimed at disrupting the inference.

Trend signal

This does not yet fit a pattern New Runtime tracks. The repository presents a focused privacy utility and attack evaluation, not evidence of a broader shift in agent architecture, workflows, or interoperability.

Why now

The repository frames the tool as a response to a demonstrated location leak: an aircraft's Doppler signature can survive common processing and remain matchable against public flight data. Its defense makes that threat more actionable by offering both a scrubber and a reference scorer for checking the result.

Connections to other events

The repository credits Colsto's demonstration of geolocating recordings from aircraft Doppler audio and ADS-B data, and says the tool is a defense plus an independent replication. That connects directly to New Runtime's earlier coverage of the attack: an anonymous tree video located through faint aircraft audio. This project shifts the practical question from whether the leak exists to whether a published copy still exposes a usable curve.

Why it matters

For teams publishing field recordings, interviews, demos, or user-generated clips, privacy review usually centers on visible metadata and identifiable speech. This project shows that ambient audio can carry location information even when a person cannot hear the relevant aircraft. Its useful design choice is pairing a transformation with an attacker-oriented check, so teams can test whether the signal remains matchable rather than treating noise reduction as proof of privacy.

The tool's limits matter to deployment: it protects only the processed copy, and the original file, live streams, calls, and CCTV are outside scope. It can also leave aircraft energy where it overlaps loud foreground such as speech harmonics. A safe workflow therefore needs to preserve the original privately, scrub the export, inspect metadata, and assess the transformed copy with known location and flight data before publication. It complements, rather than replaces, reducing long continuous recording segments.

What happens next

  • Independent testing will probe the scrubber against learned detectors and more varied recording conditions. (6 months, 60%). Watch for: reproducible evaluations using multiple devices, aircraft types, and overlapping flights.

Limits and uncertainty

The repository says the method has not been tested against learned detectors or multiple flights with a locked offset. Its calibration numbers come from a handful of synthetic cases, and borderline scorer results need care: a real Cessna flyover scored only 2–3 times chance while ranking the true spot first. Aircraft energy overlapping loud foreground can remain because the foreground is passed through; the README identifies louder flyovers under continuous speech as the most likely leak.

The reported results are the project's own tests, not an independent evaluation. They do not establish performance across different phones, recording conditions, aircraft, editing pipelines, or adversarial processing. Audio is scrubbed independently by stereo channel, which can decorrelate background noise, and quiet details near the noise floor may soften. The author characterizes the results as evidence, not a guarantee.

Skill to build now

If your team publishes location-sensitive recordings, test the exact export rather than trusting a generic denoiser. For a recording you control, capture nearby flights while recording, scrub a copy, then run the audit with the known location and flight log. If editing has changed the file's timestamp, pass the original start time explicitly. Keep the unmodified original access-controlled; the tool protects only the copy that passes through it.

Reader takeaway

Aircraft sound can leak a recording's location through a signal that is effectively inaudible, so privacy checks should account for ambient audio as well as metadata. flyover-scrub offers a concrete scrub-and-audit workflow, but its limited tests do not make the output safe by default. Treat it as one measured layer in a publishing pipeline, not a guarantee against geolocation.

Open question

How reliably does the scrubber hold up against learned detectors and recordings with multiple overlapping flights?

Evidence / sources

  1. 1github.com/lizTheDeveloper/flyover-scrub
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicSecurityExplore the security topic hub.
  2. 02topicToolsExplore the tools topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract