Well, wouldn‘t you just have to try to recognize e.g. diacritic letters (there is an according ICU property) by a “classical” character recognition model trained with non-diacritic letters? Even the general task of finding similar looking characters should easily be implemented (one might search for a model or algorithm tested with the MPEG 7 dataset). The character images should be available from the Unicode website.
Update: Hmm. After some re-thinking I guess some other tool for recognizing similar shapes might fail. Also if you would really like to compare all characters, you need to compare 100000 images every second to finish after 3 days. But for recognizing a faking of latin characters, the first idea could work (you might try to recognize all characters).
Update 2: …But in addition, beware of combining characters.