Research map / Adversarial robustness and security / Black-box and transfer attacks
An Information Theoretic Evaluation Metric for Strong Unlearning
Dongjae Jeon, Wonje Jeung, Taeheon Kim, Albert No, Jonghyun Choi
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 |
|---|---|---|---|
| evaluate unlearning with IDI in addition to black-box MIA/JSD | head distillation (HD) on CIFAR-10 ResNet-18 · single-class forgetting | black-box vs white-box efficacy: HD MIA 2.05 but IDI 1.000 | Improved |
| retrain a new head with 2% of data on frozen unlearned encoders (ablation) | RL, SALUN, Bad-T unlearned encoders · CIFAR-10 ResNet-18, forget test accuracy | forget test accuracy: Retrain no more than 41% vs over 82% for Bad-T, SALUN, RL | Hurt |
| use COLA (supervised contrastive collapse on retain set, then cross-entropy alignment) instead of RL/SALUN | class-wise forgetting · CIFAR-10 ResNet-18 | IDI: 0.830 (RL) -> 0.010 | Improved |
| use COLA instead of SCRUB | class-wise forgetting · ImageNet-1K ResNet-50 | IDI: 0.517 -> 0.040 | Improved |
| use COLA instead of Retrain | class-wise forgetting · ImageNet-1K ResNet-50 | runtime (min): 2661.90 -> 171.44 | Improved |
| use COLA instead of faster methods like HD/EU-5/FT | class-wise forgetting · CIFAR-10 ResNet-18 | runtime (min): 1.54 (EU-5) -> 4.91 | Hurt |
| use COLA+ for random data forgetting instead of SALUN | random data forgetting · CIFAR-10 ResNet-18 | IDI: 0.861 -> 0.024 | Improved |
| use IDI instead of white-box MIA to evaluate unlearning | evaluation metric · CIFAR-100 single-class forgetting ResNet-18 | metric stability: Random model MIA 53.20 vs Retrain 52.87; IDI -2.955 | Improved |
| forget set data close to Retrain MIA does not imply low IDI (CF-5) | CF-5 · ImageNet-1K five classes ResNet-50 | MIA vs IDI: MIA 10.21 vs 9.41 Retrain, IDI 0.701 | Mixed |
Closest papers by meaning
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.88)
- When to Forget? Complexity Trade-offs in Machine Unlearning (ICML 2025 · similarity 0.88)
- The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning Evaluation (NeurIPS 2026 · similarity 0.88)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.87)
- Partially Blinded Unlearning: Class Unlearning for Deep Networks from Bayesian Perspective (AAAI 2025 · similarity 0.87)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.87)
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.
