Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning