Towards Robust Detection of Adversarial Examples