Does Training with Synthetic Data Truly Protect Privacy?

TL;DR

This paper critically examines whether training with synthetic data genuinely safeguards privacy, highlighting that empirical methods lack rigorous guarantees and may falsely suggest privacy protection.

Contribution

The paper analyzes four training paradigms using synthetic data, revealing significant differences and emphasizing the need for rigorous privacy evaluation.

Findings

Different synthetic data methods vary greatly in privacy implications

Empirical privacy claims can be misleading without formal guarantees

Rigorous evaluation is essential for trustworthy privacy protection

Abstract

As synthetic data becomes increasingly popular in machine learning tasks, numerous methods--without formal differential privacy guarantees--use synthetic data for training. These methods often claim, either explicitly or implicitly, to protect the privacy of the original training data. In this work, we explore four different training paradigms: coreset selection, dataset distillation, data-free knowledge distillation, and synthetic data generated from diffusion models. While all these methods utilize synthetic data for training, they lead to vastly different conclusions regarding privacy preservation. We caution that empirical approaches to preserving data privacy require careful and rigorous evaluation; otherwise, they risk providing a false sense of privacy.