Do not use gaussian blur or pixelize to protect privacy

Today I've seen it one too many times. People using gaussian blur on screenshots to "protect privacy".

But, you say, it's heat flow simulation. It's proven to be nonreversible.

That is true in principle if your starting point is just any image. But it is completely false if your image is known to consist of, say, black text on a white background:



Let's gauss this with 20 px radius.



If you enlarge the result, you can see how easy it is to notice the bounding box of the gaussian blur. Simply find the first pixel that differs from pure white.



How does this help recover the plaintext? Above is the left part of the first letter "T". If the first letter is "A" instead, the part in question would look like this.



It is fairly easy to compare all the pixels in the image with the blurred letters of the alphabet, and with standard techniques from mathematical optimization ("minimize the total brightness distance for all the pixels"), you can calculate that the shape in question is the letter "T" with 99.96% confidence.

Then subtract the blurred letter T from the image, and move to the next letter. With a bit of experimentation on step width and exact start of the original bounding box, you get this:



Now just repeat.

Tip: a similar technique exists for pixelization, although depending on pixel block size and relative layout it can be more complex to execute. It still works almost all the time unless your pixel blocks are significantly larger than your letters.

To summarize, if you really want to protect privacy in a screenshot, please, black out the entire bounding box of the text fragment in question.