Research map / Federated learning and privacy / Differential privacy
Deep Learning With Data Privacy via Residual Perturbation
Wenqi Tao, Huaming Ling, Zuoqiang Shi, Bao Wang
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 residual perturbation (ensemble of 5) to ResNet8 | ResNet8 · IDC (gamma=0.2) | membership AUC: 0.563 -> 0.495 | Improved |
| add residual perturbation (ensemble of 5) to ResNet8 | ResNet8 · IDC (gamma=0.2) | test accuracy: 0.847 -> 0.867 | Improved |
| add residual perturbation (gamma=4.5) instead of DPSGD | ResNet8 · CIFAR10 | test accuracy: 0.625 -> 0.695 | Improved |
| add residual perturbation (gamma=3.25) instead of DPSGD | ResNet8 · CIFAR100 | test accuracy: 0.288 -> 0.383 | Improved |
| add residual perturbation (gamma=2.5) instead of DPSGD | ResNet8 · MNIST | test accuracy: 0.975 -> 0.978 | Improved |
| add residual perturbation instead of non-private training | ResNet8 · MNIST | test accuracy: 0.988 -> 0.978 | Hurt |
| add residual perturbation (gamma=0.75, ensemble of 5) | ResNet8 · CIFAR10 | membership AUC: 0.757 -> 0.573 | Improved |
| add skip connections to En5ResNet8 (ablation) | En5ResNet8 residual perturbation · CIFAR10 | test accuracy: 0.653 -> 0.675 | Improved |
| add ensemble of 5 models instead of a single model (Strategy II) (ablation) | residual perturbation · IDC | test accuracy: 0.868 -> 0.872 | Improved |
| use Strategy II noise (gamma=2.0) instead of non-private training | ResNet8 · CIFAR10 | test accuracy: 66.9 -> 71.2 | Improved |
| increase noise coefficient gamma (ablation) | residual perturbation · IDC | train-test gap | Improved |
| use residual perturbation instead of DPSGD (theoretical budget) | ResNet8 · IDC | DP epsilon: 15.79 -> 1.1e3 | Hurt |
Closest papers by meaning
- Do not Let Privacy Overbill Utility: Gradient Embedding Perturbation for Private Learning (ICLR 2021 · similarity 0.90)
- Improving the Privacy and Practicality of Objective Perturbation for Differentially Private Linear Learners (NeurIPS 2023 · similarity 0.90)
- Differentially Private Stochastic Gradient Descent with Fixed-Size Minibatches: Tighter RDP Guarantees with or without Replacement (NeurIPS 2024 · similarity 0.89)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.88)
- Shifted Interpolation for Differential Privacy (ICML 2024 · similarity 0.88)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · 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.
