Adversarial Robustness of Pruned Neural Networks