Differentially Private Sequential Data Synthesis with Structured State Space Models and Diffusion Models

Tomoya Matsumoto, Takayuki Miura, Toshiki Shibahara, Masanobu Kii, Kazuki Iwahana, Osamu Saisho, Shingo Okamura

Keywords: synthetic data, differential privacy, sequential data

TL;DR: We propose a new data synthesis method combining structured state space models and diffusion models to generate sequential data with differential privacy.

Abstract: Sequential data such as electrocardiograms and electroencephalograms are being increasingly utilized, and protecting the privacy of individuals in data has become an important issue. For statistical analysis while preserving privacy, data synthesis with differential privacy (DP) has been attracting attention. However, DP synthetic data generally suffers from a decrease in quality. In this paper, we aim to achieve high-quality DP synthesis for sequential data. First, we show that previous DP sequential data synthesis has quality issues. We then propose DP structured state space diffusion (DP-SSSD), a DP sequential data synthesis method based on novel generative AI, which combines structured state space models and diffusion models. Experiments show that DP-SSSD can generate higher-quality sequential data than the previous methods under equal privacy protection strength.

Submission Number: 67