Research map / Federated learning and privacy / Differential privacy
Automatic Clipping: Differentially Private Deep Learning Made Easier and Stronger
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 |
|---|---|---|---|
| replace Abadi's clipping with automatic clipping (AUTO-S) | DP-SGD on image tasks · MNIST, FashionMNIST, CIFAR10, ImageNette, CelebA | test accuracy (%): e.g. 90.75 -> 91.08 (CelebA Smiling); 98.04 -> 98.15 (MNIST) | Helped |
| replace Abadi's clipping with AUTO-S | RoBERTa-base full fine-tuning · MNLI-m eps=3 | accuracy (%): 82.45 -> 83.22 | Helped |
| replace Abadi's clipping with AUTO-S | RoBERTa-base · QNLI eps=3 | accuracy (%): 87.42 -> 86.91 | Hurt |
| add stability constant (AUTO-S vs AUTO-V) (ablation) | automatic clipping · MNLI-m eps=3 RoBERTa-base | accuracy (%): 81.21 -> 83.22 | Helped |
| add stability constant (AUTO-S vs AUTO-V) (ablation) | automatic clipping · E2E eps=8 GPT2 | BLEU: 63.189 -> 63.600 | Helped |
| AUTO-S on RoBERTa-large SST2 | RoBERTa-large · SST2 eps=8 | accuracy (%): 93.81 -> 94.61 | Helped |
| AUTO-S full fine-tuning vs LoRA on E2E | GPT2 · E2E eps=3 | BLEU: 58.153 -> 61.340 | Helped |
| use AUTO-S with larger model | GPT2 large vs small · E2E eps=3 | BLEU: 61.340 -> 64.180 | Helped |
| AUTO-S removes need to tune clipping threshold R when using adaptive optimizers | DP-AdamW · E2E GPT2 | sensitivity to R | Helped |
Closest papers by meaning
- Differentially Private SGD Without Clipping Bias: An Error-Feedback Approach (ICLR 2024 · similarity 0.90)
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.88)
- Delving into Differentially Private Transformer (ICML 2024 · similarity 0.87)
- GeoClip: Geometry-Aware Clipping for Differentially Private SGD (NeurIPS 2025 · similarity 0.87)
- Clipping Improves Adam-Norm and AdaGrad-Norm when the Noise Is Heavy-Tailed (ICML 2025 · similarity 0.87)
- Online Sensitivity Optimization in Differentially Private Learning (AAAI 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.