Research map / Federated learning and privacy / Differential privacy
G-PATE: Scalable Differentially Private Data Generator via Private Aggregation of Teacher Discriminators
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add random projection to gradient aggregation (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.1141 -> 0.5880 | Helped |
| increase teachers from 2000 to 4000 (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.4240 -> 0.5880 | Helped |
| increase teachers from 2000 to 4000 (ablation) | G-PATE · Fashion-MNIST eps=1 | accuracy: 0.3997 -> 0.5812 | Helped |
| projection dim 10 instead of 5 (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.4638 -> 0.5880 | Helped |
| projection dim 10 instead of 20 (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.5604 -> 0.5880 | Helped |
| use 5 bins instead of 20 (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.4706 -> 0.5880 | Helped |
| use PATE aggregation instead of DPSGD with discretization and projection (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.2026 -> 0.5880 | Helped |
| use PATE aggregation instead of DPSGD with discretization and projection (ablation) | G-PATE · MNIST eps=10 | accuracy: 0.5602 -> 0.8092 | Helped |
| G-PATE vs best baseline | G-PATE · MNIST eps=10 | accuracy: 0.8092 vs 0.8066 | No clear effect |
| G-PATE vs GS-WGAN on visual quality | G-PATE · MNIST eps=10 | Inception Score: 8.59 -> 5.16 | Hurt |
| G-PATE vs PATE-GAN at low privacy budget | G-PATE · CelebA-Hair eps=1 | accuracy: 0.3789 -> 0.4985 | Helped |
| use data-independent privacy analysis (ablation) | G-PATE · MNIST eps=1 | accuracy: 0.5880 -> 0.5483 | Hurt |
Closest papers by meaning
- PATE-GAN: Generating Synthetic Data with Differential Privacy Guarantees (ICLR 2019 · similarity 0.91)
- Scalable Private Learning with PATE (ICLR 2018 · similarity 0.90)
- Hot PATE: Private Aggregation of Distributions for Diverse Tasks (ICLR 2026 · similarity 0.89)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.86)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.86)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.85)
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.