This is pretty interesting.
I was talking about this (again, she must be getting sick of it
) with my wife who is in the computer industry, and I was explaining how the sheer number of sets with varying degrees of challenge in picking out real vs fake automatically made it a difficult problem.
Her take was that it would be easier than I think, if it was done with a decent image library, some machine learning, and at first a person who knew the set essentially teaching the machine how to tell.
This seems to be essentially the same thing, perhaps without the human teacher, but working on one set that isn't too big, and that's also pretty consistent, with few if any small differences between examples of the same card. Unlike say T206, where each front comes in at probably at least 6 distinctly different versions.
I've been down on previous ideas for this, but I think this one could work.