@inproceedings{zhao-etal-2026-cleansing,
title = "Cleansing Fine-Tuning Data to Mitigate Backdoors in {LLM}s via the Base Model{'}s Inherent Generative Abilities",
author = "Zhao, Qi and
Elster, Yan-Keon and
Wressnegger, Christian",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.248/",
pages = "4866--4886",
ISBN = "979-8-89176-444-6",
abstract = "Relying on unverified, third-party data for training large language models or outsourcing fine-tuning to untrusted parties poses a severe risk to the model{'}s integrity, giving rise to so-called data-poisoning attacks that introduce backdoors in LLMs. In this paper, we investigate to which extent fine-tuning can be hardened by using the base model{'}s intrinsic generative capabilities. We sketch a simple defense strategy using mere prompt engineering that{---}in contrast to prior work{---}neither requires knowledge of the exact backdoor trig- ger used nor clean data as a reference. Instead, we leverage the assumption that (a) potential attackers aim for unsafe outputs and (b) triggers induce language incoherence. Across downstream tasks and attack types, we show that such a simple strategy can mitigate backdoors more effectively than existing defenses. In comparison to post-hoc defenses, our method substantially reduces the attack success rate up to 85{\%} with Llama-3 while consistently preserving a high natural accuracy."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zhao-etal-2026-cleansing">
<titleInfo>
<title>Cleansing Fine-Tuning Data to Mitigate Backdoors in LLMs via the Base Model’s Inherent Generative Abilities</title>
</titleInfo>
<name type="personal">
<namePart type="given">Qi</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yan-Keon</namePart>
<namePart type="family">Elster</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christian</namePart>
<namePart type="family">Wressnegger</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Relying on unverified, third-party data for training large language models or outsourcing fine-tuning to untrusted parties poses a severe risk to the model’s integrity, giving rise to so-called data-poisoning attacks that introduce backdoors in LLMs. In this paper, we investigate to which extent fine-tuning can be hardened by using the base model’s intrinsic generative capabilities. We sketch a simple defense strategy using mere prompt engineering that—in contrast to prior work—neither requires knowledge of the exact backdoor trig- ger used nor clean data as a reference. Instead, we leverage the assumption that (a) potential attackers aim for unsafe outputs and (b) triggers induce language incoherence. Across downstream tasks and attack types, we show that such a simple strategy can mitigate backdoors more effectively than existing defenses. In comparison to post-hoc defenses, our method substantially reduces the attack success rate up to 85% with Llama-3 while consistently preserving a high natural accuracy.</abstract>
<identifier type="citekey">zhao-etal-2026-cleansing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.248/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>4866</start>
<end>4886</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Cleansing Fine-Tuning Data to Mitigate Backdoors in LLMs via the Base Model’s Inherent Generative Abilities
%A Zhao, Qi
%A Elster, Yan-Keon
%A Wressnegger, Christian
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F zhao-etal-2026-cleansing
%X Relying on unverified, third-party data for training large language models or outsourcing fine-tuning to untrusted parties poses a severe risk to the model’s integrity, giving rise to so-called data-poisoning attacks that introduce backdoors in LLMs. In this paper, we investigate to which extent fine-tuning can be hardened by using the base model’s intrinsic generative capabilities. We sketch a simple defense strategy using mere prompt engineering that—in contrast to prior work—neither requires knowledge of the exact backdoor trig- ger used nor clean data as a reference. Instead, we leverage the assumption that (a) potential attackers aim for unsafe outputs and (b) triggers induce language incoherence. Across downstream tasks and attack types, we show that such a simple strategy can mitigate backdoors more effectively than existing defenses. In comparison to post-hoc defenses, our method substantially reduces the attack success rate up to 85% with Llama-3 while consistently preserving a high natural accuracy.
%U https://aclanthology.org/2026.findings-aacl.248/
%P 4866-4886
Markdown (Informal)
[Cleansing Fine-Tuning Data to Mitigate Backdoors in LLMs via the Base Model’s Inherent Generative Abilities](https://aclanthology.org/2026.findings-aacl.248/) (Zhao et al., Findings 2026)
ACL