Research map / Continual learning / Continual fine-tuning of LLMs
PeriUn: Enhancing Unlearning by Selectively Forgetting Peripheral Samples
Hee Bin Yoo, Dong-Sig Han, Jaein Kim, Byoung-Tak Zhang
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 |
|---|---|---|---|
| unlearn only low-confidence peripheral samples of the forget set | Saliency Unlearn · TinyImageNet / VGG16-bn | ToW: 74.01 -> 96.05 | Improved |
| unlearn only peripheral samples | Saliency Unlearn · TinyImageNet / ResNet18 | ToW: 89.39 -> 97.55 | Improved |
| unlearn only peripheral samples | Saliency Unlearn · CIFAR100 / ResNet50 | ToW: 94.58 -> 97.78 | Improved |
| unlearn only peripheral samples | Saliency Unlearn · CIFAR10 / ResNet18 | ToW: 96.77 -> 97.92 | Improved |
| unlearn only peripheral samples | Saliency Unlearn · CIFAR10 / ResNet18 | confidence MIA gap: 8.53 -> 2.73 | Improved |
| unlearn only peripheral samples | Saliency Unlearn · TinyImageNet / ResNet18 | confidence MIA gap: 33.18 -> 12.34 | Improved |
| reduce selection ratio from 1.0 to 0.5 (ablation) | SU+PeriUn · TinyImageNet / VGG16-bn | ToW: 74.01 -> 96.05 | Improved |
| too small selection ratio alpha=0.1 (ablation) | SU+PeriUn · TinyImageNet / ResNet18 | ToW: 89.39 -> 80.59 | Hurt |
| too small selection ratio alpha=0.1 (ablation) | SU+PeriUn · CIFAR100 / ResNet50 | ToW: 94.58 -> 84.14 | Hurt |
| apply PeriUn to Random Label | RL · CIFAR100 / ResNet50 | ToW: 59.33 -> 95.65 | Improved |
| apply PeriUn to GA | GA · TinyImageNet / ResNet18 | ToW: 64.14 -> 64.11 | No clear effect |
| apply PeriUn to GAGD | GAGD · CIFAR10 / ResNet18 | ToW: 97.12 -> 96.64 | Mixed |
Closest papers by meaning
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.89)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · similarity 0.89)
- Fast Machine Unlearning without Retraining through Selective Synaptic Dampening (AAAI 2024 · similarity 0.89)
- Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning (NeurIPS 2026 · similarity 0.88)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 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.
