Research map / Federated learning and privacy / Differential privacy
Improving Noise Efficiency in Privacy-preserving Dataset Distillation
Runkai Zheng, Vishnu Asutosh Dasu, Yinong Oliver Wang, Haohan Wang, Fernando De La Torre
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 |
|---|---|---|---|
| add DOS (decoupled sampling and optimization) to the baseline (ablation) | Dosser · CIFAR-10 (IPC=10, eps=1) | accuracy: 41.7 -> 47.7 | Improved |
| add SER (subspace projection) to the baseline (ablation) | Dosser · CIFAR-10 (IPC=10, eps=1) | accuracy: 41.7 -> 46.5 | Improved |
| add SER on top of DOS (ablation) | Dosser · CIFAR-10 (IPC=10, eps=1) | accuracy: 47.7 -> 50.6 | Improved |
| add SER when privacy noise is small (eps=10) (ablation) | Dosser · CIFAR-10 (IPC=50, eps=10) | accuracy: 57.7 -> 57.6 | No clear effect |
| add SER on top of DOS at eps=10, IPC=50 (ablation) | Dosser · CIFAR-10 (IPC=50, eps=10) | accuracy: 61.1 -> 61.0 | No clear effect |
| increase optimization iterations beyond sampling iterations (DOS) (ablation) | Dosser · CIFAR-10 | accuracy | Improved |
| vary auxiliary dataset size in SER (ablation) | Dosser · CIFAR-10 | accuracy | No clear effect |
| use full 2048-dim PCA projection instead of no SER (ablation) | Dosser · CIFAR-10 | accuracy | Improved |
| add PEA augmentation to Dosser | Dosser · CIFAR-10 IPC=10 (eps=1) | accuracy: 44.2 -> 50.6 | Improved |
| add PEA augmentation to Dosser | Dosser · FashionMNIST IPC=10 (eps=1) | accuracy: 81.6 -> 80.1 | Hurt |
| replace NDPDC with Dosser (DOS+SER) | NDPDC · CIFAR-10 IPC=10 (eps=1) | accuracy: 39.4 -> 44.2 | Improved |
| run DP-PCA per iteration instead of fixed auxiliary-data projection | Dosser · CIFAR-10 | accuracy | Hurt |
Closest papers by meaning
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.90)
- Optimal Differentially Private Model Training with Public Data (ICML 2024 · similarity 0.89)
- Private Evolution Converges (NeurIPS 2025 · similarity 0.89)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.89)
- Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.88)
- Differentially Private Synthetic Data via Foundation Model APIs 1: Images (ICLR 2024 · similarity 0.88)
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.
