Source: www.schneier.com – Author: Bruce Schneier
Comments
Winter •
I assume that pixelation is chosen to give an impression of a text. That is, the fact that it is a string of characters with a given length.
So, the prudent way to do it is to first generate a random character string of the same length and then pixelate that string.
Or just replace it with Lorem ipsum.
Btw, the same approach might be successful with badly pixelated faces in video.
Depixelation seem to be a well studied art in Japanese Adult Video. Occasionally, I hear about people getting arrested for it.
echo •
Someone did some research for photos the police or media redacted and discovered that squinting your eyes would reveal an accurate enough impression of the face for the person to be recognisable. Experiment suggested the best minimum block size for pixilisation of faces and I’ve heard of no reports of anyone making much sense of this although I suspect there might be cases where identification of a person might be made especially if it’s from CCTV.
I suspect @Winter’s suggestions are the best approach for text. Static photos may get by with minimum block size. I don’t know about moving images if the intent is to maintain a level of aesthetic integrity with the whole image. Maybe introducing randomness would help.
Car number plates are often pixelated in the media especially of VIP’s. I wonder if this essay might cause a review of best practice.
The maths and science of it is beyond me but it’s interesting reading how cosmologists have squeezed signal out of noise or feint data. It’s also funny how so many of us were so caught up in the first blurry image of a black hole like young children staring in wonder at a colourful bug on a leaf.
Conan the deconvolutionarian •
Being model based, the deconvolution is only hypothetical and is a fwiw opinion.
Morley •
I tried a de-blurring tool a while back. It worked on my screenshot program’s blur feature. Gotta actually remove the data!
Peter •
Just a matter of time until an ai model can read pixelated text just as well as captcha.
madge •
@Winter,
Yeah, I was thinking the same thing about replacing the text. Saying “never, ever, ever” is unreasonable; this could be a fun way to mess with people, and perhaps a honeypot.
As for the pixelation of pornographic images and videos, I think that’s basically “malicious compliance”. Japanese law apparently requires censorship, so they do it, and if a clever person works around it (wink, wink), hey, that’s not the publisher’s fault. It makes me wonder whether pixelation in news reporting might be kind of the same thing: just enough plausible deniability to have a defamation, privacy invasion, or contempt-of-court charge dismissed.
Reversing the past •
@ALL
This is not a new issue, just a new use.
Those who have been involved with communications and signal processing whilst not exactly eating this stuff for breakfast have been munching on it seriously since the end of the 1950’s and begining of the 1960’s when solid state electronics became small enough and fast enough to make it practically useful in real time.
Most people get to hear two or three things about communications,
- Signal to noise ratio.
- Noise is random.
- Noise is Gaussian in characteristic.
Thus you get ‘Additive White Gaussian Noise'(AWGN) and an explanation such as
https://wirelesspi.com/additive-white-gaussian-noise-awgn/
The reality is AWGN is a simple model of reality that gets widely used and is for many things not a particularly good model.
Noise in reality is the ‘Root Mean Square'(RMS) of many signals.
N^2 = S1^2 + S2^2 + S3^2 + … Si^2
Where each signal S is in reality multipled by a ‘channel gain’ which is like an information carrying signal varying with time.
But also there is another fun thing to consider. Each signal consists of multiple time dispersed copies of it’s self. And gets called ‘Inter Symbol Interference'(ISI) which if used with care can improve the information capacity of a channel
https://www.tutorialspoint.com/digital_communication/digital_communication_pulse_shaping.htm
Knowing that noise is actually multiple sources of information not true randomness, tells you why you can ‘de-pixelate’ images that are either not random or the randomness can be modelled in some -usually- out of band way thus removed layer by layer.
If you think of pixelation as being a very low grade substitution cipher with at best poor chaining, as opposed to a One Time Pad. You can also see why it can be removed layer by layer.
As is seen with the ‘Crypto-Tux’ or ‘ECB-Penguin’ image where the Linux Tux Penguin gets encrypted using AES-128 in ‘Electronic Cipher Book’ simple substitution Mode
https://github.com/robertdavidgraham/ecb-penguin
The actual moral is
“Fully Determanistic algorithms have inverse algorithms.”
(Even supposed “One Way Functions” are invertible to a dictionary attack, especially if the actual input data has a very limited input set).
lurker •
Here we go again. There was shock, horror, when some people discovered that the “blacking out” function could be peeped under.
If you don’t want stuff to be seen, remove it completely with a big sharp knife, or better, don’t put it there in the first place.
Subscribe to comments on this entry
Sidebar photo of Bruce Schneier by Joe MacInnis.
Original Post URL: https://www.schneier.com/blog/archives/2024/05/unredacting-pixelated-text.html
Category & Tags: Uncategorized,redaction,steganography – Uncategorized,redaction,steganography
Views: 0
