Research map / Federated learning and privacy / Differential privacy
GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators
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 |
|---|---|---|---|
| sanitize only generator-side upstream gradients with Wasserstein loss and fixed clip C=1 (GS-WGAN) vs DP-SGD GAN | DP GAN · MNIST eps=10 | Inception Score: 4.76 -> 9.23 | Helped |
| GS-WGAN vs DP-SGD GAN | DP GAN · MNIST eps=10 | FID: 179.16 -> 61.34 | Helped |
| GS-WGAN vs best baseline DP-Merf AE | DP generator · Fashion-MNIST eps=10 | Inception Score: 3.68 -> 5.32 | Helped |
| GS-WGAN vs DP-Merf for downstream MLP | synthetic data · MNIST eps=10 | MLP accuracy: 0.63 -> 0.79 | Helped |
| GS-WGAN vs DP-Merf for downstream MLP | synthetic data · Fashion-MNIST eps=10 | MLP accuracy: 0.56 -> 0.65 | Helped |
| GS-WGAN vs DP-Merf, average over 13 classifiers | synthetic data · MNIST eps=10 | Avg accuracy: 0.57 -> 0.60 | Mixed |
| GS-WGAN vs Fed-Avg GAN | federated GAN · Federated EMNIST | FID: 218.24 -> 60.76 | Helped |
| GS-WGAN vs Fed-Avg GAN | federated GAN · Federated EMNIST | IS: 10.88 -> 11.25 | Helped |
| lower subsampling rate than 1/1000 (ablation) | GS-WGAN · MNIST | IS/FID at fixed epsilon | No clear effect |
| Wasserstein objective with gradient penalty instead of standard GAN loss (ablation) | GS-WGAN · gradient norm during training | gradient norm variance | Helped |
Closest papers by meaning
- Generalization in Generative Adversarial Networks: A Novel Perspective from Privacy Protection (NeurIPS 2019 · similarity 0.88)
- PATE-GAN: Generating Synthetic Data with Differential Privacy Guarantees (ICLR 2019 · similarity 0.88)
- Wasserstein Generative Adversarial Networks (ICML 2017 · similarity 0.88)
- PrAda-GAN: A Private Adaptive Generative Adversarial Network with Bayes Network Structure (AAAI 2026 · similarity 0.87)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.86)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 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.