Research map / Adversarial robustness and security / Attacks on text and graphs
Erase or Hide? Suppressing Spurious Unlearning Neurons for Robust Unlearning
Nakyeong Yang, Dong-Kyum Kim, Jea Kwon, Minsung Kim, Kyomin Jung, Meeyoung Cha
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 |
|---|---|---|---|
| add attribution-guided regularizer on negative influence (SSIUU) on top of GD | GD · FaithUn Llama-3.2 harmful attack p=0.1 | forget-set accuracy after attack: 48.13 -> 14.81 | Improved |
| add SSIUU regularizer to GD | GD · FaithUn Llama-3.2 benign attack | forget-set accuracy after attack: 33.33 -> 13.33 | Improved |
| add SSIUU regularizer to GD | GD · FaithUn Qwen-2.5 harmful attack p=0.1 | forget-set accuracy after attack: 27.78 -> 4.76 | Improved |
| add SSIUU regularizer to GD | GD · FaithUn Llama-3.2 retention | RS: 81.03 -> 84.70 | Improved |
| add SSIUU regularizer to GD | GD · TOFU Llama-3.2 harmful attack p=0.1 | forget-set ROUGE after attack: 45.41 -> 31.82 | Improved |
| add SSIUU regularizer to GD | GD · TOFU Llama-3.2 benign attack | forget-set ROUGE after attack: 20.98 -> 21.08 | No clear effect |
| regularization weight lambda 0.0001 in SSIUU (ablation) | SSIUU · FaithUn | robustness to attacks | Hurt |
| increase lambda to 0.001 or 0.01 (ablation) | SSIUU · FaithUn | robustness to attacks | Improved |
| increase lambda to 0.01 and above (ablation) | SSIUU · FaithUn | retention accuracy | Hurt |
| apply SSIUU instead of GD on influence variations | GD · FaithUn Llama-3.2 attribution analysis | negative influence variation | Improved |
| SSIUU instead of GD on logit-lens accuracy of golden answer | GD · FaithUn Llama-3.2 layers 18-27 | logit lens accuracy: approx 0.2 -> chance level (0.5) | Improved |
| SSIUU instead of GA for attribution stability under attack | GA · FaithUn Llama-3.2 p=0.1 | attribution correlation rho: 0.73 -> 0.99 | Improved |
Closest papers by meaning
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.92)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 2025 · similarity 0.92)
- Towards Robust Knowledge Unlearning: An Adversarial Framework for Assessing and Improving Unlearning Robustness in Large Language Models (AAAI 2025 · similarity 0.92)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.91)
- Textual Unlearning Gives a False Sense of Unlearning (ICML 2025 · similarity 0.91)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.90)
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.
