@inproceedings{wu-etal-2025-gracefully,
title = "Gracefully Filtering Backdoor Samples for Generative Large Language Models without Retraining",
author = "Wu, Zongru and
Cheng, Pengzhou and
Fang, Lingyong and
Zhang, Zhuosheng and
Liu, Gongshen",
editor = "Rambow, Owen and
Wanner, Leo and
Apidianaki, Marianna and
Al-Khalifa, Hend and
Eugenio, Barbara Di and
Schockaert, Steven",
booktitle = "Proceedings of the 31st International Conference on Computational Linguistics",
month = jan,
year = "2025",
address = "Abu Dhabi, UAE",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.coling-main.220/",
pages = "3267--3282",
abstract = "Backdoor attacks remain significant security threats to generative large language models (LLMs). Since generative LLMs output sequences of high-dimensional token logits instead of low-dimensional classification logits, most existing backdoor defense methods designed for discriminative models like BERT are ineffective for generative LLMs. Inspired by the observed differences in learning behavior between backdoor and clean mapping in the frequency space, we transform gradients of each training sample, directly influencing parameter updates, into the frequency space. Our findings reveal a distinct separation between the gradients of backdoor and clean samples in the frequency space. Based on this phenomenon, we propose Gradient Clustering in the Frequency Space for Backdoor Sample Filtering (GraCeFul), which leverages sample-wise gradients in the frequency space to effectively identify backdoor samples without requiring retraining LLMs. Experimental results show that GraCeFul outperforms baselines significantly. Notably, GraCeFul exhibits remarkable computational efficiency, achieving nearly 100{\%} recall and F1 scores in identifying backdoor samples, reducing the average success rate of various backdoor attacks to 0{\%} with negligible drops in clean accuracy across multiple free-style question answering datasets. Additionally, GraCeFul generalizes to Llama-2 and Vicuna. The codes are publicly available at https://github.com/ZrW00/GraceFul."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-etal-2025-gracefully">
<titleInfo>
<title>Gracefully Filtering Backdoor Samples for Generative Large Language Models without Retraining</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zongru</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pengzhou</namePart>
<namePart type="family">Cheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lingyong</namePart>
<namePart type="family">Fang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhuosheng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gongshen</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-01</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 31st International Conference on Computational Linguistics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Owen</namePart>
<namePart type="family">Rambow</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Leo</namePart>
<namePart type="family">Wanner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marianna</namePart>
<namePart type="family">Apidianaki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hend</namePart>
<namePart type="family">Al-Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barbara</namePart>
<namePart type="given">Di</namePart>
<namePart type="family">Eugenio</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Steven</namePart>
<namePart type="family">Schockaert</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Abu Dhabi, UAE</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Backdoor attacks remain significant security threats to generative large language models (LLMs). Since generative LLMs output sequences of high-dimensional token logits instead of low-dimensional classification logits, most existing backdoor defense methods designed for discriminative models like BERT are ineffective for generative LLMs. Inspired by the observed differences in learning behavior between backdoor and clean mapping in the frequency space, we transform gradients of each training sample, directly influencing parameter updates, into the frequency space. Our findings reveal a distinct separation between the gradients of backdoor and clean samples in the frequency space. Based on this phenomenon, we propose Gradient Clustering in the Frequency Space for Backdoor Sample Filtering (GraCeFul), which leverages sample-wise gradients in the frequency space to effectively identify backdoor samples without requiring retraining LLMs. Experimental results show that GraCeFul outperforms baselines significantly. Notably, GraCeFul exhibits remarkable computational efficiency, achieving nearly 100% recall and F1 scores in identifying backdoor samples, reducing the average success rate of various backdoor attacks to 0% with negligible drops in clean accuracy across multiple free-style question answering datasets. Additionally, GraCeFul generalizes to Llama-2 and Vicuna. The codes are publicly available at https://github.com/ZrW00/GraceFul.</abstract>
<identifier type="citekey">wu-etal-2025-gracefully</identifier>
<location>
<url>https://aclanthology.org/2025.coling-main.220/</url>
</location>
<part>
<date>2025-01</date>
<extent unit="page">
<start>3267</start>
<end>3282</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Gracefully Filtering Backdoor Samples for Generative Large Language Models without Retraining
%A Wu, Zongru
%A Cheng, Pengzhou
%A Fang, Lingyong
%A Zhang, Zhuosheng
%A Liu, Gongshen
%Y Rambow, Owen
%Y Wanner, Leo
%Y Apidianaki, Marianna
%Y Al-Khalifa, Hend
%Y Eugenio, Barbara Di
%Y Schockaert, Steven
%S Proceedings of the 31st International Conference on Computational Linguistics
%D 2025
%8 January
%I Association for Computational Linguistics
%C Abu Dhabi, UAE
%F wu-etal-2025-gracefully
%X Backdoor attacks remain significant security threats to generative large language models (LLMs). Since generative LLMs output sequences of high-dimensional token logits instead of low-dimensional classification logits, most existing backdoor defense methods designed for discriminative models like BERT are ineffective for generative LLMs. Inspired by the observed differences in learning behavior between backdoor and clean mapping in the frequency space, we transform gradients of each training sample, directly influencing parameter updates, into the frequency space. Our findings reveal a distinct separation between the gradients of backdoor and clean samples in the frequency space. Based on this phenomenon, we propose Gradient Clustering in the Frequency Space for Backdoor Sample Filtering (GraCeFul), which leverages sample-wise gradients in the frequency space to effectively identify backdoor samples without requiring retraining LLMs. Experimental results show that GraCeFul outperforms baselines significantly. Notably, GraCeFul exhibits remarkable computational efficiency, achieving nearly 100% recall and F1 scores in identifying backdoor samples, reducing the average success rate of various backdoor attacks to 0% with negligible drops in clean accuracy across multiple free-style question answering datasets. Additionally, GraCeFul generalizes to Llama-2 and Vicuna. The codes are publicly available at https://github.com/ZrW00/GraceFul.
%U https://aclanthology.org/2025.coling-main.220/
%P 3267-3282
Markdown (Informal)
[Gracefully Filtering Backdoor Samples for Generative Large Language Models without Retraining](https://aclanthology.org/2025.coling-main.220/) (Wu et al., COLING 2025)
ACL