Bayesian Differential Privacy for Machine Learning