Research map / Continual learning / Continual fine-tuning of LLMs
Unlearning Isn’t Forgetting: Revealing Hidden Leakage in Class Unlearning Evaluations
Ali Ebrahimpour-Boroojeny, Yian Wang, Hari Sundaram
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 |
|---|---|---|---|
| match retrained model behavior with tilted reweighting instead of existing unlearning objectives | class unlearning of CIFAR-10 ResNet18 · CIFAR-10 class unlearning | CMIA accuracy: SCAR 71.93 -> TREW 95.82 (Retrain 95.27) | Improved |
| use TREW on class unlearning | fine-tuning unlearning · CIFAR-10 U-LiRA (ResNet18) | U-LiRA accuracy (%): FT 72.57 -> TREW 71.12; TREW-2R 67.72 | Improved |
| keep standard MIA score high while ignoring class geometry (existing methods) | existing unlearning methods such as SalUn and SCRUB · CIFAR-10 class unlearning | CMIA accuracy: SalUn MIA 96.33, CMIA 9.62 vs Retrain 95.27 | Hurt |
| use TREW instead of fine-tuning with plain forget label reassignment | FT baseline · CIFAR-10 class unlearning | average gap to Retrain: -5.80 -> -0.59 | Improved |
| restrict TREW updates to two random layers (TREW-2R) (ablation) | TREW · CIFAR-10 class unlearning (ResNet18) | average gap to Retrain: -0.59 -> -1.09 | No clear effect |
| use class-level similarity scores instead of sample-dependent similarity (ablation) | TREW · class unlearning | match to Retrain | Improved |
| use positive tilt beta rather than none, avoiding extreme values (ablation) | TREW · class unlearning | closeness to Retrain | Improved |
Closest papers by meaning
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.90)
- Partially Blinded Unlearning: Class Unlearning for Deep Networks from Bayesian Perspective (AAAI 2025 · similarity 0.90)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.89)
- Textual Unlearning Gives a False Sense of Unlearning (ICML 2025 · similarity 0.89)
- Adversarial Machine Unlearning (ICLR 2025 · similarity 0.88)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.88)
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.
