Uploaded June 2026 | Updated September 2026, 3 weeks ago
Generating High-Quality Tabular Synthetic Data at Scale
Shripad Gade, Meta Platforms, Inc.
Differentially Private Synthetic Data Generation (DP-SDG) enables privacy-compliant access to sensitive tabular data by creating artificial datasets that preserve statistical properties while introducing calibrated noise. While research often focuses on straightforward scenarios, deploying DP-SDG at scale introduces significant real-world challenges.
A major challenge is Scalability and Quality, as state-of-the-art algorithms struggle with the high-dimensional data common in industry. We introduce the GEM+ algorithm, which scales SDG to industry-sized datasets with hundreds of columns within tractable runtimes, achieving a 10% improvement in accuracy over the current state-of-the-art AIM algorithm, which is known to scale poorly.
A second challenge is managing Public-Private Input Data Splits, where only a subset of columns are considered sensitive. Industry applications often involve datasets that have a public/private columns mix. We propose a framework to adapt DP-SDG methods to this vertical data split, allowing for judicious use of the differential privacy budget. Furthermore, we introduce conditional generation for both PGM-based and Generator Neural Network-based SDG, where synthetic private data is conditioned on public data, substantially improving synthetic data quality.
Authors: Samuel Maddock, Shripad Gade, Graham Cormode, Will Bullock
View the full PEPR '26 program at usenix.org/conference/pepr26/program
Generating High-Quality Tabular Synthetic Data at Scale
Shripad Gade, Meta Platforms, Inc.
Differentially Private Synthetic Data Generation (DP-SDG) enables privacy-compliant access to sensitive tabular data by creating artificial datasets that preserve statistical properties while introducing calibrated noise. While research often focuses on straightforward scenarios, deploying DP-SDG at scale introduces significant real-world challenges.
A major challenge is Scalability and Quality, as state-of-the-art algorithms struggle with the high-dimensional data common in industry. We introduce the GEM+ algorithm, which scales SDG to industry-sized datasets with hundreds of columns within tractable runtimes, achieving a 10% improvement in accuracy over the current state-of-the-art AIM algorithm, which is known to scale poorly.
A second challenge is managing Public-Private Input Data Splits, where only a subset of columns are considered sensitive. Industry applications often involve datasets that have a public/private columns mix. We propose a framework to adapt DP-SDG methods to this vertical data split, allowing for judicious use of the differential privacy budget. Furthermore, we introduce conditional generation for both PGM-based and Generator Neural Network-based SDG, where synthetic private data is conditioned on public data, substantially improving synthetic data quality.
Authors: Samuel Maddock, Shripad Gade, Graham Cormode, Will Bullock
View the full PEPR '26 program at usenix.org/conference/pepr26/program
