Research map / Federated learning and privacy / Differential privacy
Bypassing the Ambient Dimension: Private SGD with Gradient Subspace Identification
Yingxue Zhou, Steven Wu, Arindam Banerjee
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 |
|---|---|---|---|
| Project noisy gradients onto the top gradient eigenspace estimated from public data instead of using unprojected DP-SGD | DP-SGD · MNIST, ε ≤ 0.42 | test accuracy | Improved |
| Project noisy gradients onto the top gradient eigenspace estimated from public data instead of using unprojected DP-SGD | DP-SGD · Fashion MNIST, ε ≤ 0.72 | test accuracy | Improved |
| Project noisy gradients onto the top gradient eigenspace estimated from public data instead of using unprojected DP-SGD | DP-SGD · MNIST and Fashion MNIST, large ε | accuracy | Hurt |
| Replace the public-data gradient eigenspace with a Gaussian random projector (ablation) | PDP-SGD · MNIST and Fashion MNIST | accuracy | No clear effect |
| Use a projection dimension of 10 rather than the larger tested dimensions (ablation) | PDP-SGD · MNIST, ε = 0.23 | accuracy | Hurt |
| Increase the projection dimension among the tested values to 50 (ablation) | PDP-SGD · MNIST, ε = 0.23 | accuracy | Improved |
| Increase public sample size from 50 toward 150 (ablation) | PDP-SGD · MNIST, ε = 0.23 | accuracy | Improved |
| Increase public sample size from 100 to 150 (ablation) | PDP-SGD · MNIST, ε = 0.23 | accuracy | No clear effect |
| Start projection at the 15th epoch instead of the 1st epoch (ablation) | PDP-SGD and RPDP-SGD · Fashion MNIST | accuracy | Improved |
| Update the eigenspace less frequently than every iteration (ablation) | PDP-SGD · MNIST and Fashion MNIST | accuracy | Hurt |
Closest papers by meaning
- PE-SGD: Differentially Private Deep Learning via Evolution of Gradient Subspace for Text (ICLR 2026 · similarity 0.91)
- An Improved Privacy and Utility Analysis of Differentially Private SGD with Bounded Domain and Smooth Losses (AAAI 2026 · similarity 0.89)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.89)
- Nearly Tight Black-Box Auditing of Differentially Private Machine Learning (NeurIPS 2024 · similarity 0.89)
- Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach (ICLR 2024 · similarity 0.88)
- The importance of feature preprocessing for differentially private linear optimization (ICLR 2024 · similarity 0.87)
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.
