Research map / Federated learning and privacy / Differential privacy
Differentially Private Sum-Product Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes differentially private sum-product networks (DPSPNs) that learn structure and parameters under epsilon-DP so a single model can both classify and generate DP synthetic data.
Problem
Differentially private models incur privacy cost per release, and DP synthetic data generators typically lose utility, so a single private model that supports both discriminative and generative tasks is needed.
Approach
Learns an SPN top-down with differentially private instance splitting (DP k-means with K=2), data-independent random binary variable splitting, and DP histogram leaves, with a class-root sum node for classification; post-processing then allows sampling synthetic data and inference without extra privacy cost.
Key results
- Across 50 training runs DPSPN shows stable AUROC on all five classification datasets, whereas MST, PrivBayes and PATE-GAN show more variance.
- At epsilon=0.1 DPSPN-generated data gives adult AUROC 0.86 vs 0.77 for MST and 0.78 for PrivBayes, and bank AUROC 0.65 vs 0.55 for MST.
- On density estimation, DPSPN log-likelihood approaches the non-private SPN as epsilon grows, e.g. kdd -2.61 at epsilon=10 vs -2.58 for SPN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · cancer (ε=1) | AUROC: 0.82 -> 0.83 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · cancer (ε=1) | AUPRC: 0.41 -> 0.46 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · diabetes (ε=1) | AUROC: 0.63 -> 0.66 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · diabetes (ε=1) | AUPRC: 0.34 -> 0.38 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · german-credit (ε=1) | AUROC: 0.64 -> 0.65 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · german-credit (ε=1) | AUPRC: 0.39 -> 0.4 | Helped |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · adult (ε=1) | AUROC: 0.88 -> 0.87 | Hurt |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · adult (ε=1) | AUPRC: 0.71 -> 0.4 | Hurt |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · bank (ε=1) | AUROC: 0.71 -> 0.71 | No clear effect |
| Generate training data with DPSPN instead of DP synthetic-data baselines | Predictive models trained on synthetic data · bank (ε=1) | AUPRC: 0.7 -> 0.7 | No clear effect |
| Increase privacy parameter ε from 0.1 to 10 (ablation) | DPSPN density estimation · kdd | log-likelihood: -2.72 -> -2.61 | Helped |
| Increase privacy parameter ε from 0.1 to 10 (ablation) | DPSPN density estimation · nltcs | log-likelihood: -6.93 -> -6.4 | Helped |
Surprising findings
- Binary random column splits only improved synthetic-data AUROC for diabetes and german-credit, suggesting instance splitting matters more than variable splitting choice.
- For epsilon above 1, DPSPN AUROC on some datasets reached or slightly surpassed the non-private baseline.
Limitations
- Only pure epsilon-DP; approximate (epsilon, delta) extension left to future work.
- MST restricted to discrete data so comparisons on continuous datasets are partial.
- Density estimation evaluated only on binary datasets and gap to non-private SPN remains large at low epsilon.
Experimental setup
Datasets: cervical cancer german-credit diabetes bank adult kdd nltcs tretail kosarek accidents pumsb star bnetflix
Benchmarks: cervical cancer german-credit diabetes bank adult binary density estimation datasets
Metrics: AUROC AUPRC log-likelihood marginal and joint fidelity metrics
Baselines: MST PATE-GAN PrivBayes MSPN CryptoSPN
Builds on: sum-product networks LearnSPN MSPN PATE-GAN Laplace mechanism
Closest papers by meaning
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.88)
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.88)
- A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets (AAAI 2025 · similarity 0.88)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.88)
- Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.88)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.87)
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.