Distributed devices such as mobile phones can produce and store large amounts\nof data that can enhance machine learning models; however, this data may\ncontain private information specific to the data owner that prevents the\nrelease of the data. We wish to reduce the correlation between user-specific\nprivate information and data while maintaining the useful information. Rather\nthan learning a large model to achieve privatization from end to end, we\nintroduce a decoupling of the creation of a latent representation and the\nprivatization of data that allows user-specific privatization to occur in a\ndistributed setting with limited computation and minimal disturbance on the\nutility of the data. We leverage a Variational Autoencoder (VAE) to create a\ncompact latent representation of the data; however, the VAE remains fixed for\nall devices and all possible private labels. We then train a small generative\nfilter to perturb the latent representation based on individual preferences\nregarding the private and utility information. The small filter is trained by\nutilizing a GAN-type robust optimization that can take place on a distributed\ndevice. We conduct experiments on three popular datasets: MNIST, UCI-Adult, and\nCelebA, and give a thorough evaluation including visualizing the geometry of\nthe latent embeddings and estimating the empirical mutual information to show\nthe effectiveness of our approach.\n