Research map / Federated learning and privacy / Privacy attacks and secure learning
Overlearning Reveals Sensitive Attributes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the uncensored task representation to infer sensitive attributes | Models trained for the original task · UTKFace, FaceScrub, Places365, Twitter, Yelp, Health, and PIPA | sensitive-attribute accuracy over majority-class guessing | Helped |
| Apply adversarial censoring to the representation | Original-task models · UTKFace | sensitive-attribute accuracy: 62.18 -> 53.28 | Hurt |
| Apply adversarial censoring to the representation | Original-task models · FaceScrub | sensitive-attribute accuracy: 33.65 -> 30.23 | Hurt |
| Apply adversarial censoring to the representation | Original-task models · Places365 | sensitive-attribute accuracy: 31.03 -> 12.56 | Hurt |
| Apply adversarial censoring to the representation | Original-task models · Twitter | sensitive-attribute accuracy: 38.46 -> 34.27 | Hurt |
| Apply adversarial censoring to the representation | Original-task models · Yelp | sensitive-attribute accuracy: 33.09 -> 27.32 | Hurt |
| Apply information-theoretical censoring to the representation | Original-task models · FaceScrub | sensitive-attribute accuracy: 33.65 -> 10.61 | Hurt |
| Apply information-theoretical censoring to the representation | Original-task models · PIPA | sensitive-attribute accuracy: 87.95 -> 82.02 | Hurt |
| De-censor the adversarially censored representation | Representation inference attack · UTKFace | sensitive-attribute accuracy: +6.10 | Helped |
| De-censor the adversarially censored representation | Representation inference attack · FaceScrub | sensitive-attribute accuracy: +12.24 | Helped |
| Train the UTKFace gender classifier using only one race | UTKFace gender classifier · UTKFace test images across four races | race inference accuracy: 61.95%, 61.99%, 60.85% and 60.81% | No clear effect |
| Increase adversarial censoring strength γ (ablation) | FaceScrub and Twitter models · FaceScrub and Twitter | sensitive-attribute inference accuracy | No clear effect |
Closest papers by meaning
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.85)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.85)
- Generative Unlearning for Any Identity (CVPR 2024 · similarity 0.85)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.85)
- Adversarial Mixup Unlearning (ICLR 2025 · similarity 0.85)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · 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.