Beyond the Privacy-Utility Tradeoff - Differential Privacy in Tabular Data Synthesis

Explore advanced techniques for generating differentially private synthetic data that maintains high utility for AI applications while providing mathematical privacy guarantees. Learn how to leverage Gretel Safe Synthetics (now part of NVIDIA) to create synthetic datasets that preserve statistical properties and downstream task performance across heterogeneous data types. Discover strategies for calibrating privacy parameters ε and δ specifically for mixed text and tabular data, maintaining high utility on classification tasks under stringent privacy constraints with less than 0.05 difference in AUC when using differential privacy. Examine two practical use cases involving e-commerce reviews and doctor's notes to understand real-world applications of privacy-preserving data synthesis. Gain insights into quantifying resilience against membership inference and attribute inference attacks, providing a comprehensive framework for balancing privacy protection with data utility in sensitive AI applications.