PAR-GAN: Improving the Generalization of Generative Adversarial Networks Against Membership Inference Attacks.