Research: Evaluating Synthetic Data Generation Methods for Privacy-Preserving Machine L...

Research question: How do different synthetic data generation techniques impact the trade-off between data privacy and analytical utility in machine learning pipelines?

Background & Motivation: As organizations increasingly rely on data-driven decision-making, concerns over personal data privacy have intensified, driven by regulations such as GDPR and CCPA. Synthetic data generation has emerged as a promising technique to enable privacy-preserving analytics by creating artificial datasets that mimic the statistical properties of real data.

Research Gap: While various synthetic data generation methods exist—including differential privacy, generative adversarial networks (GANs), and variational autoencoders—there is limited comparative research assessing their efficacy in balancing privacy preservation with analytical utility in realistic machine learning contexts.

Approach & Expected Contribution: This project will systematically review and benchmark multiple synthetic data generation methods using standard datasets (e.g., UCI Adult, MIMIC-III). It will evaluate privacy leakage (e.g., membership inference attacks) and utility (e.g., predictive accuracy, statistical similarity) across downstream machine learning tasks. The work aims to provide a rigorous, reproducible framework for practitioners to assess privacy-utility trade-offs.

Why it matters: The findings will inform both academic research and industry practice, helping data scientists and policymakers choose appropriate synthetic data approaches for privacy-preserving analytics without sacrificing essential data-driven insights.