Robust Machine Learning via Privacy/Rate-Distortion Theory