> Perhaps specific phrases that are only likely to appear in military documents, or a classification level in a header/footer, or even some specific combination of font, font weight, line spacing and indentation.

Do you have any idea how difficult it would be to maintain a database of distinguishing sensitive marks on documents?

Hell, some documents are so protected even the labeling is protected information.

Infeasible.

Edit: I don't mean infeasible or difficult from a technical standpoint. I mean procedurally, this isn't viable. You couldn't accomplish it in a way that wouldn't be so full of holes as to be functionally useless.