Research map / Federated learning and privacy / Differential privacy
Private Set Generation with Discriminative Information
Dingfan Chen, Raouf Kerkouche, Mario Fritz
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 |
|---|---|---|---|
| constrain synthetic samples to a DCGAN generator output (image prior) (ablation) | Private Set Generation · MNIST eps=10 spc=10 | ConvNet test accuracy (%): 94.9 -> 92.2 | Hurt |
| add DCGAN generator prior (ablation) | Private Set Generation · FashionMNIST eps=10 spc=20 | ConvNet test accuracy (%): 77.7 -> 70.7 | Hurt |
| add DCGAN generator prior with 1 sample per class (ablation) | Private Set Generation · MNIST eps=10 spc=1 | ConvNet test accuracy (%): 81.4 -> 88.2 | Improved |
| add DCGAN generator prior with 1 sample per class (ablation) | Private Set Generation · FashionMNIST eps=10 spc=1 | ConvNet test accuracy (%): 66.7 -> 63.0 | Hurt |
| increase samples per class from 20 to 50 (ablation) | Private Set Generation · MNIST eps=10 | ConvNet test accuracy (%): 95.6 -> 94.0 | Hurt |
| increase samples per class from 20 to 50 (ablation) | Private Set Generation · FashionMNIST eps=10 | ConvNet test accuracy (%): 77.7 -> 71.3 | Hurt |
| increase samples per class from 1 to 10 (ablation) | Private Set Generation · MNIST eps=10 | ConvNet test accuracy (%): 81.4 -> 94.9 | Improved |
| replace DP generative models with directly optimized private synthetic set | DP-Merf · MNIST eps=10 | ConvNet test accuracy (%): 85.7 -> 95.6 | Improved |
| replace DP-Merf with directly optimized private synthetic set | DP-Merf · FashionMNIST eps=1 | ConvNet test accuracy (%): 61.2 -> 70.2 | Improved |
| transfer synthetic set from ConvNet to MLP downstream classifier | Private Set Generation · FashionMNIST eps=10 | MLP test accuracy (%): 62.2 vs 73.1 for DP-Merf | Hurt |
| add differential privacy (eps=10) to set generation | Private Set Generation · CIFAR-10 spc=20 | ConvNet test accuracy (%): 52.6 -> 42.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GS-WGAN — GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators (NeurIPS 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Private Evolution Converges (NeurIPS 2025 · similarity 0.91)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.90)
- A Privacy-Preserving Framework for Generative Model-driven Synthetic Datasets (AAAI 2025 · similarity 0.89)
- Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion (ICML 2025 · similarity 0.89)
- Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.89)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (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.
