Can Synthetic Data Reduce Conservatism of Distributionally Robust Adversarial Training?

We employ distributionally robust optimization to prevent overfitting in adversarial training and use synthetic data to reduce its conservatism.