Generative Models with Information-Theoretic Protection Against Membership Inference Attacks