Towards Interpretable Deep Neural Networks by Leveraging Adversarial Examples