Robustness Reprogramming for Representation Learning