Research map / Federated learning and privacy / Differential privacy
Differentially Private Image Classification by Learning Priors from Random Processes
Xinyu Tang, Ashwinee Panda, Vikash Sehwag, Prateek Mittal
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 Phase I synthetic-data prior and Phase II linear probing to DP-SGD | DP-SGD WRN-16-4 · CIFAR10 eps=1 | test accuracy (%): 56.8 -> 72.32 | Improved |
| Add DP-RandP phases | DP-SGD · CIFAR100 eps=3 | test accuracy (%): 30.73 -> 43.33 | Improved |
| Add DP-RandP phases | DP-SGD · DermaMNIST val eps=1 | validation accuracy (%): 69.00 -> 71.78 | Improved |
| Remove Phase II linear probing (all budget to Phase III) (ablation) | DP-RandP · CIFAR10 eps=1 | test accuracy (%): 72.32 -> 69.03 | Hurt |
| Remove Phase II (ablation) | DP-RandP · CIFAR10 eps=8 | test accuracy (%): 85.26 -> 84.45 | Hurt |
| Allocate all privacy budget to Phase II (linear probing only) (ablation) | DP-RandP · CIFAR10 eps=3 | test accuracy (%): 79.99 -> 71.08 | Hurt |
| Vary epochs of Phase II from 0 to 94 at eps=2 (ablation) | DP-RandP · CIFAR10 eps=2 | test accuracy (%): 75.30 -> 77.35 -> 63.10 | Mixed |
| Vary Phase II epochs 4,8,12,16 at eps=3 (ablation) | DP-RandP · CIFAR10 eps=3 | test accuracy (%): 79.71 / 79.74 / 79.48 / 80.27 | No clear effect |
| Use MoCo instead of alignment/uniformity loss in Phase I (ablation) | DP-RandP · CIFAR10 eps=8 | test accuracy (%): 85.26 -> 82.84 | Hurt |
| Use MoCo instead of alignment/uniformity loss in Phase I (ablation) | DP-RandP · CIFAR10 eps=1 | test accuracy (%): 72.32 -> 72.79 | No clear effect |
| Use WRN-40-4 instead of WRN-16-4 (ablation) | DP-RandP · CIFAR10 eps=1 | test accuracy (%): 72.32 -> 73.15 | Improved |
| Different synthetic datasets for Phase I (Dead leaves vs StyleGAN-Oriented) (ablation) | private linear probing · CIFAR10 eps=1 | test accuracy (%): 59.92 -> 67.79 | Improved |
Closest papers by meaning
- The importance of feature preprocessing for differentially private linear optimization (ICLR 2024 · similarity 0.91)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.91)
- Differentially Private Prototypes for Imbalanced Transfer Learning (AAAI 2025 · similarity 0.91)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.89)
- Optimal Differentially Private Model Training with Public Data (ICML 2024 · similarity 0.89)
- Differentially private optimization for non-decomposable objective functions (ICLR 2025 · 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.
