Research map / Federated learning and privacy / Privacy attacks and secure learning
Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights
Xingli Fang, Jung-Eun Kim
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 CWRF on top of HAMP | ResNet18 · CIFAR-10 | test accuracy (%): 77.79 -> 81.43 | Improved |
| add CWRF on top of CCL | ResNet18 · CIFAR-10 LiRA | attack AUC (%): 83.95 -> 64.82 | Improved |
| add CWRF on top of RelaxLoss | ResNet18 · CIFAR-10 LiRA | attack AUC (%): 70.51 -> 68.31 | Improved |
| add CWRF on top of RelaxLoss | ResNet18 · CIFAR-10 RMIA | TPR at 0.1% FPR (%): 0.52 -> 1.22 | Hurt |
| add CWRF on top of DP-SGD | ViT · CIFAR-10 | test accuracy (%): 57.63 -> 60.45 | Improved |
| add CWRF on top of RelaxLoss | ViT · CIFAR-10 LiRA | attack AUC (%): 73.45 -> 55.88 | Improved |
| add CWRF on top of HAMP | ViT · CIFAR-10 LiRA | attack AUC (%): 50.53 -> 50.15 | No clear effect |
| add CWRF on top of PAST | ViT · CIFAR-10 | test accuracy (%): 54.84 -> 54.66 | No clear effect |
| fine-tune privacy-invulnerable weights (A3) instead of privacy-vulnerable weights (A2) after rewinding (ablation) | CWRF with RelaxLoss, ResNet18, CIFAR-100 · 3.0% rewound | test cross-entropy loss: 0.9288 -> 0.8044 | Improved |
| rewind privacy-vulnerable weights (A3) instead of removing them (A1) (ablation) | CWRF with RelaxLoss, ResNet18, CIFAR-100 · CIFAR-100 | accuracy recovery | Improved |
| rewind 0.1% of weights and fine-tune invulnerable ones (A3) versus training from scratch (ablation) | RelaxLoss · test cross-entropy loss | test cross-entropy loss: 1.5398 -> 0.7720 | Improved |
| increase rewound portion from 0.1% to 5.0% (A3) (ablation) | CWRF A3 · test cross-entropy loss | test cross-entropy loss: 0.7720 -> 0.8330 | Hurt |
Closest papers by meaning
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 2025 · similarity 0.88)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.88)
- Mitigating Privacy Risk in Membership Inference by Convex-Concave Loss (ICML 2024 · similarity 0.87)
- POLAR-Bench: A Diagnostic Benchmark for Privacy-Utility Trade-offs in LLM Agents (NeurIPS 2026 · similarity 0.86)
- Learning Robust and Privacy-Preserving Representations via Information Theory (AAAI 2025 · similarity 0.86)
- Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models (NeurIPS 2024 · similarity 0.85)
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.
