AI Paper Atlas

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

ICLR 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: Adversarial prompts for language models

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 changedWhereMeasuredEffect
add attribution-guided regularizer on negative influence (SSIUU) on top of GDGD · FaithUn Llama-3.2 harmful attack p=0.1forget-set accuracy after attack: 48.13 -> 14.81Improved
add SSIUU regularizer to GDGD · FaithUn Llama-3.2 benign attackforget-set accuracy after attack: 33.33 -> 13.33Improved
add SSIUU regularizer to GDGD · FaithUn Qwen-2.5 harmful attack p=0.1forget-set accuracy after attack: 27.78 -> 4.76Improved
add SSIUU regularizer to GDGD · FaithUn Llama-3.2 retentionRS: 81.03 -> 84.70Improved
add SSIUU regularizer to GDGD · TOFU Llama-3.2 harmful attack p=0.1forget-set ROUGE after attack: 45.41 -> 31.82Improved
add SSIUU regularizer to GDGD · TOFU Llama-3.2 benign attackforget-set ROUGE after attack: 20.98 -> 21.08No clear effect
regularization weight lambda 0.0001 in SSIUU (ablation)SSIUU · FaithUnrobustness to attacksHurt
increase lambda to 0.001 or 0.01 (ablation)SSIUU · FaithUnrobustness to attacksImproved
increase lambda to 0.01 and above (ablation)SSIUU · FaithUnretention accuracyHurt
apply SSIUU instead of GD on influence variationsGD · FaithUn Llama-3.2 attribution analysisnegative influence variationImproved
SSIUU instead of GD on logit-lens accuracy of golden answerGD · FaithUn Llama-3.2 layers 18-27logit lens accuracy: approx 0.2 -> chance level (0.5)Improved
SSIUU instead of GA for attribution stability under attackGA · FaithUn Llama-3.2 p=0.1attribution correlation rho: 0.73 -> 0.99Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy