Lipika Ramaswamy

Production AI systems improve through a data flywheel: teams create training examples from curated source material, those examples shape model behavior, production usage reveals what the model still needs, and those usage signals become the next round of improvement.

As organizations increasingly leverage sensitive data for AI applications, generating high-quality synthetic data with mathematical guarantees of privacy has become crucial. This talk explores the use of Gretel Safe Synthetics (now part of NVIDIA) to generate differentially private synthetic data that maintains high fidelity to the source data and high utility on downstream tasks across heterogeneous datasets. Our analysis presents a framework for privacy-preserving synthetic data generation with two use cases: e-commerce reviews and doctor’s notes. We reveal nuanced strategies for: - Calibrating privacy parameters ε and δ for mixed text and tabular data - Maintaining statistical properties and high utility on downstream classification tasks under stringent privacy constraints (e.g. - Quantifying resilience to membership inference and attribute inference attacks