tl;dr Planes that fly overhead are audible in any audio you post online. If the time of the recording is known, then by cross-referencing with public flight data, it is possible to find the location of the video within 100 meters.

Let this be a reminder to be very careful about posting videos and images that you personally recorded. Nothing is private on the internet

you are viewing a single comment's thread
view the rest of the comments
[–] 13 points 4 days ago* (last edited 4 days ago) (7 children)

You can watch the video they discuss this a bit, but I don't think so.

Removing the airplane doppler curves from the recording seems very difficult, you cannot just easily filter them out. *

They discuss some mitigations, namely cutting the recording and adding fake airplane sounds. But cutting wouldn't work for a live stream anyway, plus the whole point of cutting would be to desync two planes, but if you already know the exact time (since it's live) one single plane might be good enough to get a fix.

Adding fake airplane sounds doesn't remove the real airplane sounds, so one could still find the real position if one ignored the fake airplanes, which could be done by assuming any possible combination of airplanes to be either fake or real and then doing the method. If the fake airplanes were random, those wouldn't correlate with anything, so if you were wise to the fact that fake random airplanes appear in the recording, you could just focus on the pairs of planes that correspond to real flight data and dismiss the rest, though that would add some computational complexity.

If you were to not just insert random flights, but rather specific flights to make it look like you were at a different location, that would still not remove the real airplane sounds from the recording, and someone who was wise to the fact you're trying to fool them could pick up on that. Plus that seem complex and would require your recording software to have full live flight data to simulate your fake location.

* Thinking about it some more, though I'm not knowledgeable enough tbh: While you couldn't easily filter this out I don't think with like a filter bank, if you did like a sliding window (which would need to be many seconds since it would need to cover the whole time the airplane is recorded), and then ran a fancy doppler curve detector on the whole window, you could maybe fully delete the doppler curve (and harmonics) in frequency space and replace it with synthetic noise without fucking up the whole recording too much.

  • source
  • parent
  • hideshow 7 child comments
  • [–] 6 points 4 days ago (5 children)

    could you remove the infrasound entirely?

  • source
  • parent
  • hideshow 5 child comments
  • [–] 8 points 3 days ago (4 children)

    They go over it in the video, you can't just filter the frequencies without either affecting normal voice frequencies and/or leaving it in there but even more silent but still detectable. Though I did update my previous post and maybe something fancy could be done with Fourier transforms but I honestly don't know enough about signal processing to be certain.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 5 points 3 days ago (3 children)

    sanest solution is deepfake btw

  • source
  • parent
  • hideshow 3 child comments
  • [–] 6 points 3 days ago (2 children)

    I just do text, airplane vibrations aren't going to click my keyboard switches.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 5 points 3 days ago (1 child)

    it’s just transformers do stuff that you have to break your brain for days with wavelets (it being audio would be sanest approach), you just synthesize 1 million text2speech recordings augmented with doppler signatures and not in all possible combinations and go to town, i suspect it will take 3 days to bang up offline filter/good enough faker this way

  • source
  • parent
  • hideshow 1 child comment
  • [–] 6 points 3 days ago* (last edited 3 days ago)

    I looked at this for a while trying to decipher this. I guess you're saying you could train, somehow, by synthetic example, a filter that would remove the doppler curves. I don't know shit about this sort of thing, that's fucking magic to me.

    Edit: When you said sanest solution was deepfake, I thought you meant people should use fake AI voices instead of recording their own voice so as not to inadvertently record the airplanes, I guess you meant it could help with filter out the doppler curves.

  • source
  • parent
  • [–] 4 points 3 days ago

    If you were to not just insert random flights, but rather specific flights to make it look like you were at a different location, that would still not remove the real airplane sounds from the recording, and someone who was wise to the fact you're trying to fool them could pick up on that. Plus that seem complex and would require your recording software to have full live flight data to simulate your fake location.

    Record the audio being played at a different location with another device (ie recording the original audio playing from the PC with your phone). I'm guessing it would be pretty easy to tell that it would be a recording of a recording, but I wonder how that would affect the Doppler curves. If the recording's Doppler curves are indistinguishable from the rerecording's Doppler curves, then it would take a lot more data crunching. Or how about record a video, record a second video a week later with the same exact script, and splice the two videos together. Maybe throw in a few fake Doppler curves to further confuse things.

    If you are someone state actors want to liquidate asap, then it's only a matter of patient number crunching. If you spliced together 3 different recordings with 7 Doppler curves and threw in 4 fake Doppler curves for a total of 11 curves, then they would number crunch by going:

    Assuming 1 recording, 0 fakes.
    Assuming 1 recording, 1 fake. Check all combinations.
    ...
    Assuming 1 recording, 10 fakes. Check all combinations.
    Assuming 2 recordings, 0 fakes. Check all combinations.
    Assuming 2 recordings, 1 fake. Check all combinations.
    ...
    Assuming 2 recordings, 9 fakes. Check all combinations.
    Assuming 3 recordings, 0 fakes. Check all combinations.
    ...
    Assuming 11 recordings, 0 fakes.

  • source
  • parent