ICML'20 paper "Confidence-Calibrated Adversarial Training: Generalizing to Unseen Attacks". - View it on GitHub