Research map / Federated learning and privacy / Differential privacy
Do not Let Privacy Overbill Utility: Gradient Embedding Perturbation for Private Learning
Da Yu, Huishuai Zhang, Wei Chen, Tie-Yan Liu
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 |
|---|---|---|---|
| replace isotropic perturbation of the full gradient with embedding and residual perturbation | GP · CIFAR-10 test, ε=8 | accuracy (%): +18.5% | Improved |
| replace isotropic perturbation of the full gradient with embedding and residual perturbation | GP · SVHN test, ε=8 | accuracy (%): +3.5 | Improved |
| replace isotropic perturbation of the full gradient with embedding and residual perturbation | GP · MNIST test, ε=8 | accuracy (%): +1.2 | Improved |
| add perturbation and release of the residual gradient (ablation) | B-GEP · CIFAR-10 test, ε=8 | accuracy (%): 63.0 -> 74.9 | Improved |
| add perturbation and release of the residual gradient (ablation) | B-GEP · SVHN test, ε=8 | accuracy (%): 92.3 -> 95.1 | Improved |
| add perturbation and release of the residual gradient (ablation) | B-GEP · CIFAR-10 validation, ε=0.1, pretrained features | accuracy (%): 91.0 -> 90.9 | No clear effect |
| use randomly labeled rather than correctly labeled CIFAR10 auxiliary samples (ablation) | GEP · CIFAR-10 test, (8,10^{-5})-DP | accuracy (%): 72.9 -> 75.1 | Improved |
| use randomly labeled CIFAR100 rather than randomly labeled CIFAR10 auxiliary samples (ablation) | GEP · CIFAR-10 test, (8,10^{-5})-DP | accuracy (%): 75.1 -> 74.7 | No clear effect |
| increase anchor-gradient count from m=1000 to m=4000 (ablation) | GEP · CIFAR-10 test, (8,10^{-5})-DP | accuracy (%): 74.6 -> 75.2 | Improved |
Closest papers by meaning
- Deep Learning With Data Privacy via Residual Perturbation (TPAMI 2025 · similarity 0.90)
- Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners (NeurIPS 2023 · similarity 0.90)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.89)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.88)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.88)
- Differentially private optimization for non-decomposable objective functions (ICLR 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.
