Unredacter: Never use pixelation as a redaction technique

So, if you want to redact a rectangle between pixels [ 12, 4 ] – [ 34, 18 ], redact [ 8, 0 ] – [ 40, 24 ] instead, this way the black rectangle will cover complete blocks.

I have also seen people try to redact by drawing a black block on a new layer in Photoshop, for example, and then exporting as PDF. They're unaware that the layers are being exported in the PDF and the block can be removed. Always flatten your document completely before export.

Whether that's enough to decode the redacted information will depend on how many pixels were redacted and other intricacies of JPEG.

This is not too dissimilar to compression oracle attacks, such as CRIME (https://en.wikipedia.org/wiki/CRIME), an infamous vulnerability in HTTPS.

Consider the following case: You have limited previously knowledge of the unredacted document; for example, you know that it was a pure black-and-white document that was compressed as a standard color JPEG (not too unusual). Suppose further that in the block extending from (16, 16) to (24, 24) (exclusive), you wish to redact the single pixel at 23, 23, presumedly as the upper left corner of a larger black bar. Finally, suppose that the post-redacted image is losslessly compressed; this simplifies the argument, but is not strictly necessary.

In the way we've phrased this problem, there are only two values for the original color of the redacted pixel; it was black, or white. The rest of the pixels in the block are known in their encoded form (unredacted), but also known in their pre-encoded form (pure black or pure white, while JPEG compression will put them somewhere into dark gray or light gray, respectively). With all known pixels set to their pre-encoded form, each possible value of the redacted pixel can be considered, using (known, deduced, or guessed) JPEG compression parameters to regenerate the compressed block. Whichever value of the redacted pixel leads to the observed encoded values of the unredacted pixels is the correct value of the redacted pixel.

(In practice, the problem might be less constrained; and you might have multiple but less than a block of pixels redacted; these make it harder to make use of the leaked information, but do not change that some information has indeed leaked.)

Programs that appear to black out a section of image, but actually leak information.

There are AI algorithms that 'enhance' blurred faces, however, the output 'looks' like real human face, but in all likelihood, different from the ground truth (i.e. the real face that got blurred). Therefore caution must be exercised in advocating such algorithms (e.g. in court of law)

Whereas this algorithm is 'just' enumerating the text, generating blur and finding the closest match to the original blur.

I doubt the US TLAs have it, because the software they have is often rubbish generated by a cartel of low skill providers, and frequently 10 years behind the times. But I bet the Chinese Ministry of State Security has it. So so many dissidents to keep track of, to identify after they have moved overseas and apply pressure to.

E.g. a page to do fold recognition on, or even to figure out what software (e.g. Gmail) and to look up what font the software uses.

If you aren't doing this so often that you need automation, you could sidestep that issue by just taking screenshots at 400% zoom or so and accurately measure how many pixels each letter (a-z) takes, as well as how many pixels a space is (the censored part might be into the spacing on each side of the word), and measure how many pixels the gap is between the words surrounding the censored part, then

    for word in wordlist do
      # Start by accounting for the spaces
      wordsize = charwidths[' '] * 2
      for char in word do
        wordsize += charwidths[char]
      done
      if wordsize == gap_size then
        print("Possible word: " + word)
      endif
    done

http://smartdeblur.net/

before the above the author made a version which was free (but old link is now dead):

https://github.com/Y-Vladimir/SmartDeblur/downloads

LEO's have access to a Commercial tool called Amped Five that is said to be very good at it:

https://ampedsoftware.com/five

That's a clever idea for speeding up this process actually. Instead of guessing a-z for every position, only guess valid word completions. If it doesn't end up giving a good score or the human judges it to be nonsensical, it can fall back to a-z guessing for that word.

A bit similar to my hangman solver (https://lucgommans.nl/p/hangman-solver/), which looks for the only words still possible with the given letters already known, but simpler because you only need prefix matching.

I know you meant it for very strong blurs, and there it is not actually an advantage because you need to go through thousands of words before guessing one (instead of 26×5≃130 guesses, assuming an average word length of ~5), but yeah there you'd have no other choice.

I've also seen people try to redact sensitive information by poorly scribbling it out using the pencil tool on their phone, leaving enough parts of letters visible to guess what was originally there, or try to "black out" text using a brush that wasn't fully opaque, allowing it to be revealed by adjusting the contrast. Basically, a lot of people don't know how to safely redact stuff.

I don't think pixelation is commonly used because most people probably don't know how to use it, but my Samsung phone's built-in image editor also has a pixelation feature, so I honestly expect to see it pop up more often in the future.