How to date a recording using background electrical noise
robertheaton.comHaving spent a decade in biometrics technology consulting, which is a similar "identify whether this thing is unique, and also you might use this in court" set of technologies... I wonder what the error rates are here.
Oddly enough, court testimony for e.g. fingerprint analysis hangs on the testimony of a human expert claiming 100% certainty, rather than the error characteristics of an automated algorithm. But we DO have recorded and proven False Match and False Non-Match rates from the manufacturers, independent companies, and NIST when it comes to algorithmic techniques. This seems similar to voice comparisons, where error rates are a function of how long the sample is.
I can see fairly easily showing that there are no clear discontinuities in the hum compared to what would be expected from random splicing (though as a defense attorney I would challenge that the very people who are presenting the clip are the ones introducing a spoofable signal; biometric error rates are against RANDOM and non-adversarial presentation, spoofing detection is an entirely different beast).
However, showing uniqueness of a hum sample that is n seconds long compared to the entire continuous history of background hum would be a more rigorous analysis. I wonder if the defense team requested that given that this was a new and unproven technique.
If this is interesting to you or your kids, here's a less detailed but pretty accessible video [1] explaining how mains hum forensics work, complete with British accent.
As soon as I saw this topic I knew someone would bring the Tom Scott. Excellent work!
Thank you - I was very interested!
Also described in http://hummingbirdclock.info/
There was a IEEE Signal Processing Society student competition to do this in 2016.[1] I've been trying to find the results of the competition, but the relevant links are dead. The Internet Archive may be able to help.
It takes hundreds of seconds of data to see power line frequency changes on a major grid. All that synchronous rotating machinery attached to the grid has to physically change speed slightly, and there's huge inertia. Here's data for the UK national grid.[2] (This is supposed to be live, but is 2 days old.) It looks like the control systems are set up to take action at 0.1Hz error, because the frequency wanders around in that range, but as soon as it gets outside, there's a speedup or slowdown to get it back in bounds.
Claims made about short audio samples are probably bogus.
[1] https://sigport.org/sites/default/files/Information%20on%20t...
I have occasionally wondered if a tape recording (audio, VHS, etc) would capture something of the earths magnetic field at the time and position of the information was committed to the media, and if this could be discerned.
It's not useful for mag tape or wire recordings, but archaeologists have available a technique that can date artifacts like campfires, pottery kilns, and burned out adobe houses within the last 10,000 years.
It's called archaeomagnetic dating[0]. It turns out that heated ferromagnetic materials, such as magnetite, capture the magnitude and direction of the earth's magnetic field as they cool down through the Curie temperature[1]. That allows an investigator to ascertain the direction of magnetic north the last time a likely sample was heated above the Curie point. Over time, magnetic north changes with respect to true north.
[0]https://en.wikipedia.org/wiki/Paleomagnetism [1]https://en.wikipedia.org/wiki/Curie_temperature
While reading it I was wondering if it was possible to do it with videos too and the lights containing the mains "flickering", since videos has lower sampling rates than audio it must be harder. It looks like it's harder but possible and there is an improved method for it in this paper:
Spain to Turkey seems like quite a distance in a synchronous grid. Is it delayed and warped if we compare the hums in two different locations?
It doesn't answer your question at all but you might find this interesting: https://hackaday.com/2018/03/09/europe-loses-six-minutes-due...
I think technically speaking there could be a delay between two points. However, the "hum" being fingerprinted is the precise frequency which is tightly coupled through electromechanical action of all the generators on the grid spinning collectively at the same speed.
Im sure the waveform wouldn't be perfectly matched from one generation area with another, but when you look at a longer period of time, say a few seconds, you wouldn't be able to refute the waveform match.
Could we use a phase-locked loop instead of a sliding window FFT? I guess only if you're sure the recording contains no splices?
Yes, indirectly, by using the PLL as a demodulator to extract the sidebands from the 50/60 Hz "carrier" frequency. These types of problems usually boil down to reducing the signal bandwidth as far as you can in order to get rid of as much noise as possible, and the loop filter in a PLL can be good for that.
Some terms to Google for more information on that would be "synchronous detection" and "lock-in amplifier."
The method in the article talks about Fourier transform techniques, but in reality, this is a correlation problem that doesn't have to be handled in the frequency domain at all. Essentially you'd do a dot product of the contents of a sliding window from the recording against the utility's own recording of the AC power waveform. When the peak value is reached, the window offset corresponds to the best estimate of the signal's position with respect to the timeframe of the recording. This benefits tremendously from bandpass filtering, in terms of saving computation time, but doesn't strictly require it.
In real life, you'd use the STFT or something like it as the author describes, but you'd use it as a convolution filter, not to locate the frequency peak. That's kind of a red herring in an otherwise-excellent article.