On the Connection between Differential Privacy and Adversarial Robustness in Machine Learning