Research map / Federated learning and privacy / Differential privacy
SlaClip: Gradient Norm Slacks can be Indicator for Adaptive Clipping in DP-SGD
Shuyan Zou, Shaowei Wang, Zhanxing Zhu, Jin Li, Changyu Dong, Vladimiro Sassone, Han Wu
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 near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · MNIST (eps=1) | test accuracy (%): 92.05 -> 94.15 | Improved |
| add near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · IMDB (eps=2) | test accuracy (%): 73.42 -> 74.64 | Improved |
| add near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · CIFAR-10 (eps=4) | test accuracy (%): 59.52 -> 60.41 | Improved |
| add near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · F-MNIST (eps=1) | test accuracy (%): 83.87 -> 84.56 | Improved |
| add near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · MNIST (eps=3) | test accuracy (%): 97.38 -> 97.48 | No clear effect |
| add near-zero CDF signal (dynamic target ratio) to SlaClip-Q (ablation) | SlaClip · F-MNIST (eps=3) | test accuracy (%): 87.20 -> 87.24 | No clear effect |
| replace Adap-Clip with SlaClip (no extra private query) | DP-SGD · MNIST (eps=1) | test accuracy (%): 92.67 -> 94.15 | Improved |
| replace AutoClip with SlaClip | DP-SGD · IMDB (eps=2) | test accuracy (%): 75.51 -> 74.64 | Hurt |
| replace AutoClip with SlaClip | DP-SGD · CIFAR-10 (eps=4) | test accuracy (%): 60.95 -> 60.41 | Hurt |
| replace Vanilla-Clip with SlaClip | DP-SGD · IMDB (eps=2) | test accuracy (%): 61.62 -> 74.64 | Improved |
| replace Vanilla-Clip with SlaClip | DP-SGD · MNIST (eps=1) | test accuracy (%): 92.23 -> 94.15 | Improved |
| replace DC-SGD-E with SlaClip | DP-SGD · CIFAR-10 (eps=4) | test accuracy (%): 53.22 -> 60.41 | Improved |
Closest papers by meaning
- GeoClip: Geometry-Aware Clipping for Differentially Private SGD (NeurIPS 2025 · similarity 0.92)
- Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach (ICLR 2024 · similarity 0.91)
- Scalable DP-SGD: Shuffling vs. Poisson Subsampling (NeurIPS 2024 · similarity 0.89)
- Online Sensitivity Optimization in Differentially Private Learning (AAAI 2024 · similarity 0.88)
- Clipped SGD Algorithms for Performative Prediction: Tight Bounds for Stochastic Bias and Remedies (ICML 2025 · similarity 0.88)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 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.
