Research map / Federated learning and privacy / Differential privacy
Don’t Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence
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 |
|---|---|---|---|
| Use DP-Sinkhorn instead of GS-WGAN | Private image generation · MNIST | FID: 61.3 -> 48.4 | Helped |
| Use DP-Sinkhorn instead of DP-MERF | Private image generation · Fashion-MNIST | FID: 132.6 -> 128.3 | Helped |
| Use DP-Sinkhorn instead of DP-MERF | Private image generation · Fashion-MNIST | CNN accuracy: 75.4 -> 71.1 | Hurt |
| Use DP-Sinkhorn instead of DP-MERF | Private image generation · CelebA 32x32 | FID: 274.0 -> 189.5 | Helped |
| Perturb image gradients rather than parameter gradients (ablation) | DP-Sinkhorn with L2 cost and no debiasing · MNIST | FID: 218.6 -> 124.3 | Helped |
| Add L1 cost to the L2 cost (ablation) | DP-Sinkhorn with image-gradient perturbation and no debiasing · MNIST | CNN accuracy: 80.8 -> 65.7 | Hurt |
| Mix L1 and L2 costs rather than using L1 alone (ablation) | DP-Sinkhorn with image-gradient perturbation and no debiasing · MNIST | CNN accuracy: 65.7 -> 77.6 | Helped |
| Use semi-debiasing rather than no debiasing (ablation) | DP-Sinkhorn with image-gradient perturbation and L1+L2 cost · MNIST | FID: 89.1 -> 48.4 | Helped |
| Use semi-debiasing rather than full debiasing (ablation) | DP-Sinkhorn with image-gradient perturbation and L1+L2 cost · MNIST | CNN accuracy: 66.7 -> 83.2 | Helped |
| Increase semi-debiasing fraction p from 0.2 to 0.4 (ablation) | DP-Sinkhorn · Fashion-MNIST | FID: 163.4 -> 133.0 | Helped |
| Increase semi-debiasing fraction p from 0.2 to 0.4 (ablation) | DP-Sinkhorn · MNIST | FID: 74.5 -> 48.4 | Helped |
| Use BigGAN rather than DCGAN as generator (ablation) | DP-Sinkhorn with L2 cost · CelebA 32x32 | FID: 156.7 -> 168.4 | Hurt |
Closest papers by meaning
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.90)
- Differentially Private Prototypes for Imbalanced Transfer Learning (AAAI 2025 · similarity 0.88)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.87)
- Exactly Minimax-Optimal Locally Differentially Private Sampling (NeurIPS 2024 · similarity 0.86)
- RAPID: Retrieval Augmented Training of Differentially Private Diffusion Models (ICLR 2025 · similarity 0.86)
- Optimal Differentially Private Model Training with Public Data (ICML 2024 · similarity 0.86)
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.