Robust Machine Learning via Privacy/ Rate-Distortion Theory