The problem isn't only detecting suspicious statistical or computational patterns, it's doing the detection on a device you ultimately don't control, and entirely in software (at least for PCs, at least for now).
With enough effort, all software can be virtualized, and whether the defender's effort can even theoretically be scaled more easily than the attacker's is an open question.
With enough effort, all software can be virtualized, and whether the defender's effort can even theoretically be scaled more easily than the attacker's is an open question.