As ML-as-a-service gains popularity, Stanford researcher explores how one can steal learned models; modeled as a learning task, of course.