Preventing Catastrophic Overfitting in Fast Adversarial Training: A Bi-level Optimization Perspective

Open in new window