BOLT: Boolean-feature One-release Label Transfer for Differentially Private Tabular Synthetic Data
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Differentially private tabular synthesis must balance predictive utility with the computational cost of producing usable data. We introduce BOLT (Boolean-feature One-release Label Transfer), which constructs two complementary models from a single jointly calibrated Gaussian release: a simple class-conditional covariate sampler and a richer Bernoulli na\"ive-Bayes classifier over fixed threshold predicates and randomly selected two-literal conjunctions. BOLT samples covariates from the simpler model and assigns stochastic labels using the classifier, incorporating private interaction statistics into prediction without fitting a generator to reproduce those statistics. Under a fixed public schema and encoding, synthesis is postprocessing of the shared -differentially private release. We evaluate five private synthesis methods across five datasets, four privacy budgets (), and two downstream classifiers. Averaging F1 and AUROC ratios to the corresponding nonprivate real-data references equally across datasets, budgets, learners, and metrics, BOLT retains 87.0% of reference utility with a mean complete-table synthesis time of seconds, compared with 89.3% and hours for AIM. BOLT achieves the lowest mean synthesis time in all twenty dataset–privacy configurations. Ratios of recorded arithmetic mean runtimes give and speed advantages over DP-MERF, the next-fastest method, and AIM, respectively. After averaging scores across releases, BOLT exceeds each of DP-MERF, DP-HP, and Tab-PE in at least 80% of the dataset–budget–learner–metric comparisons.