DP-MERF: Differentially Private Mean Embeddings with Random Features for Practical Privacy-Preserving Data Generation

TL;DR

DP-MERF introduces a novel differentially private data generation method leveraging random feature representations of kernel mean embeddings, enabling efficient privacy-preserving synthetic data generation with improved utility.

Contribution

The paper presents a new DP data generation approach using random features for kernel mean embeddings, reducing privacy costs and eliminating hyper-parameter tuning for sensitivity.

Findings

Achieves better privacy-utility trade-offs than existing methods.

Applicable to heterogeneous tabular and image data.

Requires only a single perturbation during training.

Abstract

We propose a differentially private data generation paradigm using random feature representations of kernel mean embeddings when comparing the distribution of true data with that of synthetic data. We exploit the random feature representations for two important benefits. First, we require a minimal privacy cost for training deep generative models. This is because unlike kernel-based distance metrics that require computing the kernel matrix on all pairs of true and synthetic data points, we can detach the data-dependent term from the term solely dependent on synthetic data. Hence, we need to perturb the data-dependent term only once and then use it repeatedly during the generator training. Second, we can obtain an analytic sensitivity of the kernel mean embedding as the random features are norm bounded by construction. This removes the necessity of hyper-parameter search for a clipping norm to handle the unknown sensitivity of a generator network. We provide several variants of our algorithm, differentially-private mean embeddings with random features (DP-MERF) to jointly generate labels and input features for datasets such as heterogeneous tabular data and image data. Our algorithm achieves drastically better privacy-utility trade-offs than existing methods when tested on several datasets.