Generative Adversarial Network (GAN) and its variants have shown promising\nresults in generating synthetic data. However, the issues with GANs are: (i)\nthe learning happens around the training samples and the model often ends up\nremembering them, consequently, compromising the privacy of individual samples\n- this becomes a major concern when GANs are applied to training data including\npersonally identifiable information, (ii) the randomness in generated data -\nthere is no control over the specificity of generated samples. To address these\nissues, we propose imdpGAN - an information maximizing differentially private\nGenerative Adversarial Network. It is an end-to-end framework that\nsimultaneously achieves privacy protection and learns latent representations.\nWith experiments on MNIST dataset, we show that imdpGAN preserves the privacy\nof the individual data point, and learns latent codes to control the\nspecificity of the generated samples. We perform binary classification on digit\npairs to show the utility versus privacy trade-off. The classification accuracy\ndecreases as we increase privacy levels in the framework. We also\nexperimentally show that the training process of imdpGAN is stable but\nexperience a 10-fold time increase as compared with other GAN frameworks.\nFinally, we extend imdpGAN framework to CelebA dataset to show how the privacy\nand learned representations can be used to control the specificity of the\noutput.\n