Our work of data augmentation for adversarial training, IDBH, was accepted at ICLR2023.