Benn Jordan of The Flashbulb fame (who showed up here on HN recently for saving a PNG to a bird and then retrieving it) has a YT video where he shows off several interesting techniques for capturing audio using only video capture, which seem similar to this technique. https://youtu.be/mEC6PM97IRI
You can do something similar with a laser. The laser is reflected off of an object and into an receiver where it is converted into audio. The two techniques differ in methods, but they both rely on measuring the movement of secondary materials affected by sound waves which can then be used to infer speech.
It's interesting that measuring speech vibration is possible, but on the whole I don't see this as viable attack. One would need to get real close to a target, the cited 60% accuracy is at 50 cm distance. At 300 cm they report 2% which seems on the level of guessing words randomly.
The "remotely" in the title is no warranted in my view. Of course there may be ways to improve on their results. But so far it seems to me that a highly directional microphone would outperform the radar at any distance. One interesting aspect here is how sensitive the attack is to ambient noise. So maybe sitting close to a target with a radar hidden in headphones, one could glean some words even in a noisy environment like a subway. But I'd still put my money on a microphone performing better in that environment.
Now I wonder whether the attack would work better on headphones. They are smaller and should vibrate more. On the other hand the surface is smaller, so on the whole they might not provide more signal to a radar.
Most humans would assume that giving an app permission to use the accelerometer is not the same as giving the app permission to use the microphone. It's yet another form of permission confusion that allows for unintended privacy violations.
A microphone could work together with a radar; correlating the two signals may help tell the sound from the noise.
I suppose that at certain radio frequencies human bodies are mostly translucent, while the small metal membrane is still resolvable. When a head of a person covers well the acoustic signal (the phone is at the opposite ear), a microphone is helpless, but a radar us still fairly usable.
Of course Penn State. See Geschickter Fund (an CIA cover for MKULTRA research) and the School of International Affairs (SIA) (another CIA recruitment front).
That was likely Gyrophone (2014) from Stanford/Rafael researchers who demonstrated accelerometers and gyroscopes could capture speech frequencies at 80-250Hz - sufficient for speaker identification and some speech recognition without requiring microphone permissions.
reply