Research map / Federated learning and privacy / Differential privacy
SecretGen: Privacy Recovery on Pre-trained Models via Distribution Discrimination
Zhuowen Yuan, Fan Wu, Yunhui Long, Chaowei Xiao, Bo Li
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 cutout transformations to the discrimination metric (ablation) | SecretGen · CelebA, face.evoLVe, center mask | attack accuracy (Protocol 1): 0.528 -> 0.550 | Improved |
| add cutout transformations to the discrimination metric (ablation) | SecretGen · CelebA, face.evoLVe, face T mask | attack accuracy (Protocol 1): 0.256 -> 0.273 | Improved |
| add cutout transformations to the discrimination metric (ablation) | SecretGen · CelebA, VGG16, center mask | attack accuracy (Protocol 1): 0.584 -> 0.599 | Improved |
| add cutout transformations to the discrimination metric (ablation) | SecretGen · CelebA, ResNet152, center mask | attack accuracy (Protocol 1): 0.594 -> 0.592 | No clear effect |
| add ground truth label knowledge | SecretGen · CelebA whitebox, face.evoLVe, center mask | attack accuracy (Protocol 1): 0.551 -> 0.788 | Improved |
| add ground truth label knowledge | SecretGen · CelebA blackbox, face.evoLVe, center mask | attack accuracy (Protocol 1): 0.447 -> 0.603 | Improved |
| use SecretGen instead of GMI, both with GT label | SecretGen vs GMI · CelebA whitebox, face.evoLVe, face T mask | attack accuracy (Protocol 1): 0.467 -> 0.695 | Improved |
| replace GMI (GT label) with SecretGen without GT label | SecretGen · CelebA whitebox, face.evoLVe, center mask | attack accuracy (Protocol 1): 0.595 -> 0.551 | Hurt |
| replace GMI (GT label) with SecretGen without GT label | SecretGen · CelebA whitebox, VGG16, center mask | attack accuracy (Protocol 2): 0.955 -> 0.928 | Hurt |
| apply purification defense to the target model | SecretGen · CelebA blackbox, VGG16, center mask | attack accuracy (Protocol 1): 0.351 -> 0.328 | Hurt |
| apply purification defense to the target model | SecretGen · CelebA blackbox, VGG16, face T mask | attack accuracy (Protocol 2): 0.837 -> 0.747 | Hurt |
| replace sequential cutout with flipping, grayscale or color jitter (ablation) | pseudo label predictor · CelebA, face.evoLVe | label prediction accuracy | Hurt |
Closest papers by meaning
- Differentially Private Prototypes for Imbalanced Transfer Learning (AAAI 2025 · similarity 0.87)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 2024 · similarity 0.87)
- Does Training with Synthetic Data Truly Protect Privacy? (ICLR 2025 · similarity 0.87)
- Position: Considerations for Differentially Private Learning with Large-Scale Public Pretraining (ICML 2024 · similarity 0.86)
- Differentially Private Sum-Product Networks (ICML 2024 · similarity 0.86)
- Privacy Backdoors: Stealing Data with Corrupted Pretrained Models (ICML 2024 · 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.
