Research map / Federated learning and privacy / Differential privacy
Differentially Private Synthetic Data via APIs 4: Tabular Data
Toan Tran, Arturs Backurs, Zinan Lin, Victor Reis, Li Xiong, Sergey Yekhanin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace XOR-stress-test baselines' marginal queries with Tab-PE full-record nearest-neighbor voting | Tab-PE · XOR simulation (5 features, eps=1.0) | AUC: baselines random guess -> Tab-PE 0.8 | Improved |
| use random-walk variation with scheduled mutation decay instead of crossover and mutation operators (ablation) | Tab-PE · Artificial Characters | accuracy: 43% -> 50% | Improved |
| use two-stage selection (sampling then ranking) instead of ranking-only or sampling-only (ablation) | Tab-PE · Artificial Characters (eps=1.0) | validation accuracy and 5-WD | Improved |
| sampling-only selection (no ranking) in later iterations (ablation) | Tab-PE · Artificial Characters (eps=1.0) | downstream accuracy | Hurt |
| polynomial decay of mutation rate instead of linear decay (ablation) | Tab-PE · Artificial Characters | utility | Improved |
| increase marginal query degree for baselines (ablation) | query-based baselines (AIM, GEM, PrivGSD) · NN prior SCM dataset | accuracy and runtime | No clear effect |
| generate synthetic samples at 10-20% of the original dataset size (ablation) | Tab-PE · eps=1.0 | downstream utility | Improved |
| run Tab-PE for 15-20 iterations rather than fewer (ablation) | Tab-PE · hyperparameter sensitivity | utility | Improved |
| use Tab-PE instead of AIM on low-order real datasets | Tab-PE · widely used low-order datasets | accuracy: about 1% accuracy drop vs AIM | Hurt |
| use Tab-PE instead of AIM on Tree and RFF SCM priors | Tab-PE · SCM simulation, eps=1.0 | accuracy: roughly 10% gains over best baselines; about 4% for NN prior | Improved |
| use Tab-PE instead of AIM at low-order marginal fidelity | Tab-PE · Artificial Characters and Person Activity | 1-3 way WD | Hurt |
| use Tab-PE instead of AIM for runtime | Tab-PE · NN prior SCM, eps=1.0, CPU | running time: 28x faster than AIM, 10x faster than PrivMRF | Improved |
Closest papers by meaning
- Private Evolution Converges (NeurIPS 2025 · similarity 0.92)
- PCEvolve: Private Contrastive Evolution for Synthetic Dataset Generation via Few-Shot Private Data and Generative APIs (ICML 2025 · similarity 0.90)
- Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.90)
- CuTS: Customizable Tabular Synthetic Data Generation (ICML 2024 · similarity 0.89)
- Do You Really Need Public Data? Surrogate Public Data for Differential Privacy on Tabular Data (NeurIPS 2025 · similarity 0.89)
- HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
