A good candidate could be Benford's law [1], which makes predictions about the distribution of digits. If the digits of the numbers found in the results section of the paper deviate too much from this law, it could be a red flag.
And to be honest, once it was known to be used in this way, it would be simple to make fake data comply with the law.
Big picture, the goal is not to catch every fake. It's to (1) catch enough so that smarter fakers learn to submit their stuff elsewhere (vs. to the journal, conference, or institution in question, which really cares about fake detection). And (2) dimmer fakers are mostly caught ~promptly.