Research map / Adversarial robustness and security / Attacks on text and graphs
Membership Inference Attacks for Unseen Classes
Pratiksha Thaker, Neil Kale, Steven Wu, Virginia Smith
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 |
|---|---|---|---|
| exclude the target class from shadow model training (unseen-class setting) (ablation) | LiRA shadow-model attack · CINIC-10 and CIFAR-100 (superclasses) | TPR at 1% FPR | Hurt |
| use quantile regression instead of LiRA shadow models on unseen CIFAR-100 classes | Membership inference attack · CIFAR-100 superclasses, unseen class | TPR at 1% FPR: 1.4% -> 3.8% | Improved |
| use quantile regression instead of shadow models on CINIC-10 | Membership inference attack · CINIC-10, unseen class | TPR at 1% FPR | Improved |
| use RMIA instead of LiRA | Shadow-model attack in unseen-class setting · CIFAR-100 / CINIC-10 unseen class | TPR at low FPR | Hurt |
| use top-two logit difference as score function instead of true-label confidence (ablation) | Quantile regression attack · Class dropout | TPR at 1% FPR | Improved |
| use top-two logit difference as score function instead of true-label confidence (ablation) | Shadow-model attack · Class dropout | TPR at 1% FPR | Hurt |
| early stopping of the quantile regression model (ablation) | Quantile regression attack · CINIC-10 / CIFAR-100 | attack performance | No clear effect |
| leave out 100 to 900 of 1000 classes from attack training (ablation) | Quantile regression attack · ImageNet, unseen classes | TPR at 1% FPR: stays around 3.8% (0.4% at 0.1% FPR) | No clear effect |
| reduce attack training data to 10 samples per class (ablation) | Quantile regression attack · ImageNet | TPR at 1% FPR: 9.0% -> 3.9% | Hurt |
| use quantile regression instead of RMIA on tabular and text data | Membership inference attack · Texas and 20 Newsgroups | TPR at 1% FPR: up to 2x on Texas, up to 6x on 20 Newsgroups | Improved |
| increase regularization defense (weight decay, differential privacy) (ablation) | Target model · Unseen-class MIA | MIA success | Hurt |
Closest papers by meaning
- Causal Evaluation of Membership Inference Attacks (NeurIPS 2026 · similarity 0.88)
- Estimating Model-Level Membership Inference Vulnerability Without Reference Models (NeurIPS 2026 · similarity 0.87)
- SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing (NeurIPS 2026 · similarity 0.87)
- Sequential Membership Inference Attacks (NeurIPS 2026 · similarity 0.87)
- Assessing Per-Sample Membership Inference Vulnerability without Retraining (NeurIPS 2026 · similarity 0.86)
- Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning (NeurIPS 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.
