Research map / Federated learning and privacy / Differential privacy
Minimax Filter: Learning to Preserve Privacy from Inference Attacks
Jihun Hamm
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 minimax (adversarially trained) linear filter instead of PCA/Rand projection | linear filter, d=10-100 · GENKI | gender inference accuracy: 0.90 -> 0.5 | Improved |
| use a nonlinear minimax filter instead of non-private projections | filter on speech features · ENTERFACE | subject identification accuracy: 0.4 - 0.5 -> 0.02 | Improved |
| use minimax linear filter instead of PCA/Rand/PPLS | filter on motion features · HAR | utility-privacy compromise: 0.70 -> 0.067 | Improved |
| add minimax filtering on top of noise (Minimax-pre/post vs PCA-pre/post) (ablation) | noisy filters, d=20 · GENKI | private task accuracy without noise: 0.8 -> 0.5 | Improved |
| add minimax filtering to noisy filters (ablation) | noisy filters, d=20 · ENTERFACE | private task accuracy without noise: 0.5 -> 0.03 | Improved |
| add minimax filtering to noisy filters (ablation) | noisy filters, d=20 · HAR | private task accuracy without noise: 0.3 -> 0.07 | Improved |
| increase privacy noise level from epsilon^-1=0 to 10 (ablation) | all noisy filters · GENKI, ENTERFACE, HAR | target and private accuracy | Hurt |
| apply noise before filtering (pre) rather than after (post) (ablation) | PCA and Minimax · GENKI and HAR | target task accuracy | Improved |
| apply noise before vs after filtering (ablation) | PCA and Minimax · ENTERFACE | target task accuracy | No clear effect |
| increase filter dimension d from 10 to 100 (ablation) | Rand, PCA, PPLS · GENKI | utility-privacy tradeoff (target minus private accuracy) | No clear effect |
Closest papers by meaning
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.88)
- 𝑓-Differential Privacy Filters: Validity and Approximate Solutions (NeurIPS 2026 · similarity 0.87)
- MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective (ICML 2024 · similarity 0.87)
- Attack-Aware Noise Calibration for Differential Privacy (NeurIPS 2024 · similarity 0.87)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.86)
- Optimizing Noise Distributions for Differential Privacy (ICML 2025 · 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.
