DP-LFlow: Differentially private latent flow for scalable sensitive image generation

Dihong Jiang, Sun Sun

Abstract: Privacy concerns grow with the success of modern deep learning models, especially when the training set contains sensitive data. Differentially private generative model (DPGM) can serve as a solution to circumvent such concerns by generating data that are distributionally similar to the original data yet with differential privacy (DP) guarantees. While differentially private stochastic gradient descent (DP-SGD) is currently the leading algorithm for training a private deep learning model, the norm of the stochastic noise introduced to the model by DP-SGD increases linearly with model dimensions, which is likely to spoil the model utility (particularly so for strong DP guarantees). Among various generative models, the flow generative model is exceptional for its capability of \emph{exact} density estimation, and DP-flow in high dimensional space is rarely explored, thus it is of interest in this work. We will first show that it is challenging (or even infeasible) to train a DP-flow on image sets with acceptable utility, and then give an effective solution by reducing the generation from the pixel space to a lower dimensional latent space. We show the effectiveness and scalability of the proposed method via extensive experiments. Notably, our method is scalable to high-resolution image datasets, which is rarely studied in related works.