I do think it is unreasonable. With a SEM you only get to look at the surface of things, which is going to be either glass or metal or polysilicon. The only way to see a transistor in a sem is if you chemically remove all the top layers (which are the connections between transistors), or perform a cross section.
In the cross section case you are going to see a few dozen transistors out of the millions in a design of any complexity.
It would be remotely feasible to discover some sort of shenanigans if you knew the exact layout of the design, which would basically mean you are a foundry yourself. In that case you might get lucky and spot some difference between the mask you made, and the mask that was used to produce the part under inspection.
But the scales involved make this not believable to me. It would be roughly like scanning the whole of, say, America, and checking every street and intersection of every town, and comparing it against some known quantity to see if something changed in Springfield Missouri.
Maybe somebody could automate this, but the chemical processes for removing layers is less than perfect. Those strands of metal stretching across the ASIC have some built in tension, and if you remove the layer of glass above them, then tend to spring up and jumble. Good luck trying to do something with that.
