It could also be encoded in the choice of serif or sans-serif for each character: http://elonka.com/friedman/

That would have never occurred to me. So many tricks.

I wonder if round tripping texts could be an effective sanitizer. Text to speech and back. English to Chinese and back.

See e.g. Spread Spectrum Image Steganography for embedding messages under the noise level.

http://www.lia.deis.unibo.it/Courses/RetiDiCalcolatori/Proge...

Just have the background image have a pattern to it that changes by user: you can encode stuff in a uniform-looking background by making some squares be color (255,255,255) and others be (255,255,254) for example: it would appear to be uniform white but not really.

The classic image stegonography technique is embedding bits in the least-significant-bit of each pixel, such that the differences do not produce visual changes but the information is preserved in the raw bytes of the file. The problem with this technique is that the encoded information will not survive a screenshot (since it’s in the LSB, the colors don’t actually change, just the raw bits).

In order to survive a screenshot (or even some of the common processing steps run after uploading an image somewhere), you would need to visibly change the colors, as you say. This could work but could also produce identifiable artifacts.

(Actually, now I’ve typed this out, I’m not sure if a LSB stego image wouldn’t survive a screenshot. I guess it depends on how the screenshot function is implemented — actually would be curious if anyone has tested this.)

LSB alterations would be in lossless encoded images, like PNG (of the bit-depth was sufficient); but not in lossy image formats like JPEG.

Right, and since most screenshots are jpeg, that means LSB would likely be lost.

At least on macOS, screenshots are PNG. Does Windows make JPEGs?

The Snipping Tool included with Win 10 defaults to saving as PNG

Has been done, see my comment above.

http://failheap-challenge.com/showthread.php?12731-Once-upon...

The images are 404, but they used a background with slightly different color and something similar to a QR code.

The background trick is cool, but what happens if the user removes it entirely? Lexical Steganography might be a more suitable approach: http://web.mit.edu/keithw/tlex/

Just skimmed it, but is that not close to method #2 the same Alliance used: http://failheap-challenge.com/showthread.php?16311-Taking-th...