Robustness Guarantees for Adversarially Trained Neural Networks