Research map / Optimization / Nonconvex and smooth optimization
Beyond Uniform Lipschitz Condition in Differentially Private Optimization
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 the minimum rather than maximum per-sample Lipschitz constant as the clip norm (ablation) | DP-SGD for convex softmax-layer training · Caltech-256 | test accuracy: 50.34 ± 0.57 % -> 73.54 ± 0.32 % | Helped |
| Use the minimum rather than maximum per-sample Lipschitz constant as the clip norm (ablation) | DP-SGD for convex softmax-layer training · Food-101 | test accuracy: 37.36 ± 1.26 % -> 48.84 ± 0.89 % | Helped |
| Use the minimum rather than maximum per-sample Lipschitz constant as the clip norm (ablation) | DP-SGD for convex softmax-layer training · CIFAR-100 | test accuracy: 45.40 ± 0.96 % -> 54.40 ± 0.36 % | Helped |
| Use a smaller clip norm instead of the maximum per-sample Lipschitz constant (ablation) | DP-SGD for convex softmax-layer training · CIFAR-10 | test accuracy: 81.78 ± 0.16 % -> 83.79 ± 0.20 % | Helped |
| Use a smaller clip norm instead of the maximum per-sample Lipschitz constant (ablation) | DP-SGD for convex softmax-layer training · TweetEval Emoji | test accuracy: 25.55 ± 0.26 % -> 26.04 ± 0.29 % | Helped |
| Use a smaller clip norm instead of the maximum per-sample Lipschitz constant (ablation) | DP-SGD for convex softmax-layer training · Emotion | test accuracy: 50.34 ± 0.46 % -> 51.86 ± 0.70 % | Helped |
| Use a smaller clip norm instead of the maximum per-sample Lipschitz constant (ablation) | DP-SGD for convex softmax-layer training · EMNIST | test accuracy: 56.36 ± 1.04 % -> 62.85 ± 0.89 % | Helped |
| Use a smaller clip norm instead of the maximum per-sample Lipschitz constant (ablation) | DP-SGD for convex softmax-layer training · Fashion-MNIST | test accuracy: 79.99 ± 0.18 % -> 82.82 ± 0.18 % | Helped |
| Estimate the minimum Lipschitz constant privately with RNMM before DP-SGD (ablation) | DP-SGD · Caltech-256, total ε=2 | test accuracy: 74.03±0.47 -> 70.03±0.66 | Hurt |
| Estimate the minimum Lipschitz constant privately with RNMM before DP-SGD (ablation) | DP-SGD · Food-101, total ε=2 | test accuracy: 49.26±1.00 -> 45.84±2.00 | Hurt |
| Estimate the minimum Lipschitz constant privately with RNMM before DP-SGD (ablation) | DP-SGD · CIFAR-100, total ε=2 | test accuracy: 55.10±0.26 -> 52.89±1.02 | Hurt |
| Estimate the minimum Lipschitz constant privately with RNMM before DP-SGD (ablation) | DP-SGD · Caltech-256, total ε=4 | test accuracy: 79.00±0.04 -> 78.27±0.05 | No clear effect |
Closest papers by meaning
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.92)
- Private Stochastic Convex Optimization with Heavy Tails: Near-Optimality from Simple Reductions (NeurIPS 2024 · similarity 0.90)
- Differentially Private Optimization with Sparse Gradients (NeurIPS 2024 · similarity 0.89)
- Improved Rates of Differentially Private Nonconvex-Strongly-Concave Minimax Optimization (AAAI 2025 · similarity 0.89)
- Public-data Assisted Private Stochastic Optimization: Power and Limitations (NeurIPS 2024 · similarity 0.88)
- Differential Private Stochastic Optimization with Heavy-tailed Data: Towards Optimal Rates (AAAI 2025 · 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.