Distributional Adversarial Training utilizes cont. adv. training on diffusion-based adv. examples to close a gap in population-robust risk estimation.