On the (Statistical) Detection of Adversarial Examples