Research map / Federated learning and privacy / Privacy attacks and secure learning
Bounding the Invertibility of Privacy-preserving Instance Encoding using Fisher Information
Kiwan Maeng, Chuan Guo, Sanjay Kariyappa, G. Edward Suh
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 dFIL optimizations (compression layer, SNR regularizer) over plain noise (ablation) | dFIL split inference · CIFAR-10 ResNet-18 early, 1/dFIL=10 | accuracy: 10.70% -> 74.44% | Improved |
| add dFIL optimizations over plain noise (ablation) | dFIL split inference · CIFAR-10 ResNet-18 middle, 1/dFIL=100 | accuracy: 12.94% -> 84.27% | Improved |
| add dFIL optimizations over plain noise (ablation) | dFIL split inference · MovieLens-20M NCF-MLP, 1/dFIL=100 | AUC: 0.6120 -> 0.8081 | Improved |
| add dFIL optimizations over plain noise, split late (ablation) | dFIL split inference · GLUE-SST2 DistilBert, 1/dFIL=10 | accuracy: 90.25% -> 83.03% | Hurt |
| add dFIL optimizations over plain noise, split late (ablation) | dFIL split inference · GLUE-SST2 DistilBert, 1/dFIL=100 | accuracy: 82.68% -> 82.82% | No clear effect |
| split earlier in the network | dFIL split inference · CIFAR-10 ResNet-18, 1/dFIL=10 | accuracy: 92.35% -> 74.44% | Hurt |
| increase 1/dFIL from 10 to 100 (less privacy budget of noise) | dFIL training on encoded data · CIFAR-10, CIFAR-100 pretrained encoder | accuracy: 80.16% -> 70.27% | Hurt |
| pretrain encoder on held-out CIFAR-10 instead of CIFAR-100 | dFIL training on encoded data · CIFAR-10, 1/dFIL=100 | accuracy: 70.27% -> 78.65% | Improved |
| dFIL-based encoder instead of DP-based encoder at same unbiased MSE bound | split inference · CIFAR-10 ResNet-18, MSE bound 1e-5 | accuracy: 64.64% -> 93.09% | Improved |
| dFIL-based instead of DP-based encoder at looser bound 1e-2 | split inference · CIFAR-10 ResNet-18, MSE bound 1e-2 | accuracy: 33\% -> 87.52% | Improved |
| reduce randomized smoothing noise from 0.25 to 0.01 for score matching (ablation) | dFIL bound estimation · MNIST/CIFAR-10 | bound validity | Hurt |
| use arbitrary-attack bound (Bound-ours) instead of unbiased bound | dFIL bound · synthetic Gaussian, MNIST, CIFAR-10 | bound validity | Improved |
Closest papers by meaning
- Leveraging Per-Instance Privacy for Machine Unlearning (ICML 2025 · similarity 0.86)
- Instance-Specific Asymmetric Sensitivity in Differential Privacy (NeurIPS 2024 · similarity 0.86)
- Certification for Differentially Private Prediction in Gradient-Based Training (ICML 2025 · similarity 0.86)
- Privacy Amplification Persists under Unlimited Synthetic Data Release (NeurIPS 2026 · similarity 0.85)
- Auditing Privacy Mechanisms via Label Inference Attacks (NeurIPS 2024 · similarity 0.85)
- DECO-Bench: Unified Benchmark for Decoupled Task-Agnostic Synthetic Data Release (NeurIPS 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.
