<ul><li>Privacy attacks, specifically membership inference attacks (MIAs), are commonly used to evaluate the privacy of generative models for tabular synthetic data.</li><li>This paper highlights the importance of data domain extraction in generative models and its impact on privacy attacks.</li><li>Three strategies for defining the data domain are examined: using an externally provided domain, extracting it directly from the input data, and extracting it with differential privacy (DP) mechanisms.</li><li>The study shows that using the second approach of extracting the data domain directly from the input data can compromise end-to-end DP guarantees and make models vulnerable to privacy attacks.</li></ul>

Understanding the Impact of Data Domain Extraction on Synthetic Data Privacy

Discover more