Towards Deep Learning Models Resistant to Transfer-based Adversarial Attacks via Data-centric Robust Learning