[Cryptography] "The myth of video anonymity"
Jerry Leichter
leichter at lrw.com
Wed Dec 14 07:20:27 EST 2016
AI techniques to "de-pixelate" faces that have been deliberately pixelated for anonymity.
It's always seemed to me that this should be possible: A single pixelated image only gives you averages of pixel values over a grid (though even that might correlate surprisingly well with a database of possible values); but in a video of even a couple of seconds, there are hundreds of different somewhat overlapping averages taken, which will highly constrain the original image - especially if there's a priori information about the image (such as the fact that it's a human face).
Our brains already do something related when you walk by a picket fence getting repeated partial snapshots of what's behind it at different angles - and forming a pretty clear overall picture.
http://storagemojo.com/2016/12/13/the-myth-of-video-anonymity/
There's a link to the paper at arxiv.org - which has a cert that expired a month ago!
-- Jerry
More information about the cryptography
mailing list