The visual microphone [like laser microphone but for video]

[spacediver]

Very very cool stuff - video says it all.

Basically, they're analyzing high speed camera footage of vibrations caused by soundwaves, and using algorithms to figure out the spectral properties of those sound waves.

http://people.csail.mit.edu/mrub/VisualMic/

[Chief Blur Buster]

They even take advantage of the rolling scan nature of a camera sensor to gain more audio resolution. Very interesting, and makes you think, go "Hmmmmm."

This is like using a laser microphone to listen to a window pane vibrating, but into a video instead of real life! Spooky in a way. Amazing you can pull some audio information out of that. I wonder when we go to 8K at high bit rates, that is more data to detect audio on surfaces in video... One might be able to technologically pull speech more reliably off surfaces in a video. Spooky, spying into a video...(!)

[RealNC]

In the past they came up with a device that could capture the output of CRT monitors from a distance, purely by picking up the emissions of the CRT while it was operating.

They come up with some crazy stuff alright.

[flood]

I wonder when we go to 8K at high bit rates, that is more data to detect audio on surfaces in video... One might be able to technologically pull speech more reliably off surfaces in a video. Spooky, spying into a video...(!)

well mostly you need low noise video with high framerate. haven't looked through the paper, but I'm pretty sure the framerate basically sets a nyquist limit unless the video has rolling scan and low exposure time.

related:
the cardiio app in the apple app store can (just barely) detect your heart rate by looking at your face.

[spacediver]

In the past they came up with a device that could capture the output of CRT monitors from a distance, purely by picking up the emissions of the CRT while it was operating.

They come up with some crazy stuff alright.

I've never heard about the emissions being used (you talking about the emissions at the cathode?), but this paper ( https://www.cl.cam.ac.uk/~mgk25/ieee02-optical.pdf ) is similar.

[RealNC]

I've never heard about the emissions being used (you talking about the emissions at the cathode?), but this paper ( https://www.cl.cam.ac.uk/~mgk25/ieee02-optical.pdf ) is similar.

I can't find it anymore either. It was an article in an issue of c't (major German computer mag) somewhere around 1994 or such. They've actually built one themselves which was kind of working with very limited range. The point was "if we can do this with a $500 budget, imagine what government agencies can do."

[flood]

here's the relevant wikipedia article
http://en.wikipedia.org/wiki/Van_Eck_phreaking

lcd's are also susceptible

In April 2004, academic research revealed that flat panel and laptop displays are also vulnerable to electromagnetic eavesdropping. The required equipment for espionage was constructed in a university lab for less than US$2000

....that's quite scary

[Chief Blur Buster]

well mostly you need low noise video with high framerate. haven't looked through the paper, but I'm pretty sure the framerate basically sets a nyquist limit unless the video has rolling scan and low exposure time.

Low noise video, yes.
High frame rate, not necessary.
You can spy on audio inside a video, at a mere 60fps@60Hz with a existing low-noise off-the-shelf camera.

The pre-requisite is only framerate matching the video refresh rate, then you can let horizontal scanrate do the rest. So 60fps@60Hz video will do just fine, as long as it's low-noise. Resolution and low noise is more important than frame rate here, as long as it's a unique scanline per horizontal scanrate (no repeated refreshes).

Even NTSC video can do 15 KHz of horizontal scan rate, but to get more audio information, you want 1080p 60Hz, which is approximately 67KHz of horizontal scan rate. Allowing for oversampling, that's theoretically enough to detect up to 33KHz of audio frequency, but you want more resolution by having more vertical resolution. You would have interruptions caused by the vertical blanking interval (0.5ms-1.0ms audio dropouts every 16.7ms), but even those could be interpolated-in for intelligible speech.

Horizontal scanrate = number of scanlines per second, aka number of rows of pixels per second. Video is scanned sequentially, even in the digital era, and pushed over cables sequentially, even if packetized. You'd analyze vibrations on a per-scanline basis, rolling from top to bottom, as the next scanline is 1/(horizontal scanrate)th of a second delayed from the previous scanline above it.

You only need 60 frames per second at 60Hz refresh rate, to spy on audio in a video, by zooming your camera into a vibrating potato chip bag, so that it fills the whole frame (and the rolling-scan nature of the camera shutter & the nature of the horizontal scanrate makes it unnecessary to use high-framerate video).

[flood]

well even with rolling shutter/scan you still are limited by the exposure time. e.g. you're not going to easily see 2000hz vibrations when your exposure time is 1ms.

though it's not an exactly hard limit like nyquist because the exposure window is a square function so perhaps it's possible to extract information from the differences between rows. it's basically a time-domain superresolution problem

and if you decrease exposure time, then shot noise increases

if you look at their spectrograms, even with their several thousand hz video, it is still difficult to reproduce frequencies above 2khz.