Adversarially Robust Learning Could Leverage Computational Hardness