Towards Adversarially Robust Data-Efficient Learning with Generated Data

A well-established study of adversarial training has been proposed to improve network robustness against adversarial examples in the context of deep learning. However, its performance highly relies on large-scale training data. To relieve from such a data-hungry learning nature, we propose an efficient extension of adversarial training by conducting a data reduction method from a new perspective of generated data. The reduced dataset can be regarded as an alternative pre-training dataset, which promotes adversarial training methods for better robustness even than the original dataset. Experimental results and analyses demonstrate the effectiveness of our data reduction method, achieving the same level of adversarial robustness with a dataset volume reduced to 80% of its original size.