Research map / Federated learning and privacy / Differential privacy
Interpretable and Differentially Private Predictions
Frederik Harder, Matthias Bauer, Mijung Park
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 a stronger privacy guarantee (lower ε) for LLM | LLM · Henan Renmin Hospital Data | accuracy: 82.0±0.4% -> 79.8±0.4% | Hurt |
| Use a stronger privacy guarantee (ε=0.5 instead of ε=2) | DP-LLM · MNIST | accuracy: 94.2±0.4% -> 91.8±0.4% | Hurt |
| Use a stronger privacy guarantee (ε=2 instead of ε=0.5) | DP-LLM · Fashion-MNIST | accuracy: 83.2±0.4% -> 80.7±0.6% | Hurt |
| Add data-independent random projections (ablation) | DP-LLM · high-dimensional image data | accuracy | Improved |
| Remove random projections (ablation) | DP-LLM · image classification | interpretability | Improved |
| Increase the number of linear maps per class (ablation) | DP-LLM · MNIST and Fashion-MNIST | accuracy | Hurt |
| Increase the number of linear maps per class (ablation) | non-private LLM · MNIST and Fashion-MNIST | accuracy | Improved |
| Increase random-projection output dimensionality (ablation) | DP-LLM · MNIST and Fashion-MNIST | accuracy | Improved |
| Increase the number of linear maps per class (ablation) | DP-LLM · Henan Renmin Hospital Data | accuracy | Hurt |
| Increase the number of linear maps per class (ablation) | non-private LLM · Henan Renmin Hospital Data | accuracy | Hurt |
| Use stronger privacy (ε < 0.1) | DP-LLM · Henan Renmin Hospital Data | interpretability | Hurt |
Closest papers by meaning
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.87)
- Position: Considerations for Differentially Private Learning with Large-Scale Public Pretraining (ICML 2024 · similarity 0.85)
- Differentially Private Model Merging (NeurIPS 2026 · similarity 0.85)
- Learning from End User Data with Shuffled Differential Privacy over Kernel Densities (ICLR 2025 · similarity 0.85)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.85)
- Efficiently Computing Similarities to Private Datasets (ICLR 2024 · similarity 0.84)
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.
