Research map / Federated learning and privacy / Differential privacy
DP-KFC: Data-Free Preconditioning for Privacy-Preserving Deep Learning
Marc Molina Van den bosch, Riccardo Taiello, Albert Sund Aillet, Andrea Protani, Miguel Angel Gonzalez Ballester, Luigi Serio
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 |
|---|---|---|---|
| use synthetic-noise KFAC preconditioning instead of plain DP-SGD | DP-SGD · MNIST CNN (epsilon=1) | test accuracy (%): 91.7 -> 94.2 | Improved |
| use synthetic-noise KFAC preconditioning instead of DP-Adam | DP-Adam on BERT · StackOverflow (epsilon=1) | accuracy (%): 89.5 -> 91.8 | Improved |
| use synthetic-noise KFAC preconditioning instead of DP-SGD | DP-SGD logistic regression · IMDB (epsilon=2.8) | accuracy (%): 83.1 -> 85.8 | Improved |
| use synthetic probes instead of public data for KFAC factors | Public DP-KFC · StackOverflow BERT (epsilon=1) | accuracy (%): 96.1 -> 91.8 | Hurt |
| use synthetic pink noise instead of MNIST proxy data | Public DP-KFC · PathMNIST<-MNIST transfer (epsilon=1) | test accuracy (%): 73.4 -> 78.2 | Improved |
| use synthetic pink noise instead of MNIST proxy data | Public DP-KFC · Fashion<-MNIST transfer (epsilon=1) | test accuracy (%): 87.6 -> 87.8 | No clear effect |
| replace public proxy with pink noise in AdaDPS | AdaDPS · PathMNIST<-MNIST transfer (epsilon=1) | test accuracy (%): 70.5 -> 71.2 | No clear effect |
| add block-diagonal KFAC geometry preconditioning rather than diagonal AdaDPS | AdaDPS · MNIST CNN (epsilon=1) | test accuracy (%): 91.3 -> 94.2 | Improved |
| combine Synthetic DP-KFC with post-privatization DP-AdamBC | Synthetic DP-KFC · MNIST CNN (epsilon=1) | test accuracy (%): 94.2 -> 95.5 | Improved |
| use synthetic DP-KFC instead of DP-AdamBC alone | DP-AdamBC · MNIST CNN (epsilon=1) | test accuracy (%): 94.0 -> 94.2 | Improved |
| train frozen-backbone CrossViT with DP-KFC rather than DP-Adam | DP-Adam · CIFAR-100 CrossViT | test accuracy (%): about +1.4% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DiSK — DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction (ICLR 2025) · better on 3 of 3 shared results
- DP-AdamBC — DP-AdamBC: Your DP-Adam Is Actually DP-SGD (Unless You Apply Bias Correction) (AAAI 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Shifted Interpolation for Differential Privacy (ICML 2024 · similarity 0.88)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.88)
- DOPPLER: Differentially Private Optimizers with Low-pass Filter for Privacy Noise Reduction (NeurIPS 2024 · similarity 0.87)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.87)
- Towards hyperparameter-free optimization with differential privacy (ICLR 2025 · similarity 0.87)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · 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.
