Research map / Federated learning and privacy / Privacy attacks and secure learning
Privacy for Free: How does Dataset Condensation Help Privacy?
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 DM condensation with random rather than real-data initialization (ablation) | DM · FashionMNIST | loss-based MIA advantage (%): $83.67\pm 2.77$ -> $-0.29\pm0.48$ | Helped |
| use DM condensation with random rather than real-data initialization (ablation) | DM · CIFAR-10 | loss-based MIA advantage (%): $64.20\pm 4.77$ -> $1.06\pm 1.20$ | Helped |
| use DM condensation with random rather than real-data initialization (ablation) | DM · CelebA | loss-based MIA advantage (%): $47.00\pm 19.52$ -> $-0.56\pm1.52$ | Helped |
| use random initialization for DM-synthesized data | DM · FashionMNIST, CIFAR-10 and CelebA | loss-based MIA advantage (%): advantages are around $0$ | No clear effect |
| use DM instead of DP-MERF (ε=2) | FashionMNIST synthetic-data training · FashionMNIST, r_ipc=0.002 | accuracy (%): $60.41\pm0.78$ -> $80.59 \pm 0.62$ | Helped |
| use DM instead of DP-MERF (ε=2) | FashionMNIST synthetic-data training · FashionMNIST, r_ipc=0.01 | accuracy (%): $55.14\pm0.61$ -> $85.10 \pm 0.51$ | Helped |
| use DM instead of DP-MERF (ε=2) | FashionMNIST synthetic-data training · FashionMNIST, r_ipc=0.02 | accuracy (%): $56.39\pm 0.45$ -> $86.13 \pm 0.34$ | Helped |
| add ZCA preprocessing to KIP (ablation) | KIP · FashionMNIST, r_ipc=0.002 | accuracy (%): $73.70 \pm 1.13$ -> $74.37 \pm 0.96$ | Helped |
| add ZCA preprocessing to KIP (ablation) | KIP · FashionMNIST, r_ipc=0.01 | accuracy (%): $68.11 \pm 1.33$ -> $70.03 \pm 0.84$ | Helped |
| increase the DM synthetic-data ratio from r_ipc=0.01 to r_ipc=0.02 (ablation) | DM · FashionMNIST | accuracy (%): $85.10 \pm 0.51$ -> $86.13 \pm 0.34$ | Helped |
Closest papers by meaning
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.87)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.87)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.86)
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.86)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.86)
- Efficiently Computing Similarities to Private Datasets (ICLR 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.