On the Differential Privacy of Bayesian Inference