There's OCR systems which work on the basis of internal font consistency. They break the page into a series of single character images, and because the same character repeated is close to identical it's trivial to match them up, so you can easily build a map of characters.
You then just need a human to label each character once. With a pixel image comparison 0 looks completely different from o.
If they're using a standard font then a regular OCR (you'd only need four nines accuracy to get it 100% correct) would be fine, even with a weird font it still be easy to get that level of accuracy.
The obvious solution to this would be to cycle randomly between fonts every few characters (or keep a running total of the font used for each particular symbol, and ensure it stays below some threshold). For bonus points you could convert the key from base64/ascii to unicode or similar.
The obvious solution to this would be to cycle randomly between fonts every few characters (or keep a running total of the font used for each particular symbol, and ensure it stays below some threshold).
This sounds like a useful defence in general against OCR re-use of particular things you might publish. I wonder if it could be done in a manner unobtrusive to the eye, but progressively more expensive to algorithms, either in terms of memory or time. This is really a neat idea you have.
Doesn't matter. A few wrong characters can easily be brute-forced. Once you have enough of the characters, you can just write a program to try modifying a few of them until you get a key that works.
