This line of thought has a long research history. Check it out at http://scholar.google.com/scholar?q=fuzzing+machine+learning... .