Towards Efficient Data-Centric Robust Machine Learning with Noise-based Augmentation