DualEdit: Mitigating Safety Fallback in LLM Backdoor Editing via Affirmation-Refusal Regulation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jiang, Houcheng, Zhao, Zetong, Fang, Junfeng, Ma, Haokai, Wang, Ruipeng, Wang, Xiang, He, Xiangnan, Deng, Yang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912979508789248
author Jiang, Houcheng
Zhao, Zetong
Fang, Junfeng
Ma, Haokai
Wang, Ruipeng
Wang, Xiang
He, Xiangnan
Deng, Yang
author_facet Jiang, Houcheng
Zhao, Zetong
Fang, Junfeng
Ma, Haokai
Wang, Ruipeng
Wang, Xiang
He, Xiangnan
Deng, Yang
contents Safety-aligned large language models (LLMs) remain vulnerable to backdoor attacks. Recent model editing-based approaches enable efficient backdoor injection by directly modifying a small set of parameters to map triggers to attacker-desired behaviors. However, we find that existing editing-based attacks are often unstable under safety alignment: the edited model may start with an affirmative prefix but later revert to refusals during generation. We term this phenomenon safety fallback. To mitigate it, we propose DualEdit, a dual-objective model editing framework that simultaneously promotes affirmative tokens and suppresses refusal tokens. DualEdit further addresses two key challenges, objective imbalance and refusal diversity, via two complementary techniques: (1) dynamic loss weighting, which calibrates the relative scales of the two objectives using the pre-edited model to stabilize optimization, and (2) value anchoring, which clusters representative attention value vectors to form compact anchors, reducing conflicts from overly diverse token sets and improving generalization. Experiments on safety-aligned LLMs show that DualEdit improves attack success by 10% and reduces safety fallback rate by 11% over baselines.
format Preprint
id arxiv_https___arxiv_org_abs_2506_13285
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DualEdit: Mitigating Safety Fallback in LLM Backdoor Editing via Affirmation-Refusal Regulation
Jiang, Houcheng
Zhao, Zetong
Fang, Junfeng
Ma, Haokai
Wang, Ruipeng
Wang, Xiang
He, Xiangnan
Deng, Yang
Computation and Language
Safety-aligned large language models (LLMs) remain vulnerable to backdoor attacks. Recent model editing-based approaches enable efficient backdoor injection by directly modifying a small set of parameters to map triggers to attacker-desired behaviors. However, we find that existing editing-based attacks are often unstable under safety alignment: the edited model may start with an affirmative prefix but later revert to refusals during generation. We term this phenomenon safety fallback. To mitigate it, we propose DualEdit, a dual-objective model editing framework that simultaneously promotes affirmative tokens and suppresses refusal tokens. DualEdit further addresses two key challenges, objective imbalance and refusal diversity, via two complementary techniques: (1) dynamic loss weighting, which calibrates the relative scales of the two objectives using the pre-edited model to stabilize optimization, and (2) value anchoring, which clusters representative attention value vectors to form compact anchors, reducing conflicts from overly diverse token sets and improving generalization. Experiments on safety-aligned LLMs show that DualEdit improves attack success by 10% and reduces safety fallback rate by 11% over baselines.
title DualEdit: Mitigating Safety Fallback in LLM Backdoor Editing via Affirmation-Refusal Regulation
topic Computation and Language
url https://arxiv.org/abs/2506.13285