@inproceedings{chen-etal-2026-tuning,
title = "Tuning Just Enough: Lightweight Backdoor Attacks on Multi-Encoder Diffusion Models",
author = "Chen, Ziyuan and
Jeong, Yujin and
Braun, Tobias and
Rohrbach, Anna",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1381/",
doi = "10.18653/2026.findings-emnlp.1381",
pages = "29361--29379",
ISBN = "979-8-89176-462-0",
abstract = "As text-to-image (T2I) diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior works largely focus on models with a single text encoder, leaving recent multi-encoder architectures, e.g., Stable Diffusion 3 (SD 3), underexplored. In this paper, we study SD 3, which uses three distinct text encoders, and systematically analyze its text-encoder-based backdoor vulnerabilities. To understand each text encoder{'}s role in backdoor attacks, we define four target categories spanning from global-level to entity-level semantics and identify the minimal encoder subset required for each attack type. Furthermore, we propose Multi-Encoder Lightweight aTtacks (MELT) and validate MELT{'}s effectiveness and parameter efficiency. MELT matches full encoder fine-tuning attack performance while using fewer than (0.2{\%}) of its parameters, demonstrating that effective backdoors can be implanted under limited parameter budgets."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chen-etal-2026-tuning">
<titleInfo>
<title>Tuning Just Enough: Lightweight Backdoor Attacks on Multi-Encoder Diffusion Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ziyuan</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yujin</namePart>
<namePart type="family">Jeong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tobias</namePart>
<namePart type="family">Braun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Rohrbach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>As text-to-image (T2I) diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior works largely focus on models with a single text encoder, leaving recent multi-encoder architectures, e.g., Stable Diffusion 3 (SD 3), underexplored. In this paper, we study SD 3, which uses three distinct text encoders, and systematically analyze its text-encoder-based backdoor vulnerabilities. To understand each text encoder’s role in backdoor attacks, we define four target categories spanning from global-level to entity-level semantics and identify the minimal encoder subset required for each attack type. Furthermore, we propose Multi-Encoder Lightweight aTtacks (MELT) and validate MELT’s effectiveness and parameter efficiency. MELT matches full encoder fine-tuning attack performance while using fewer than (0.2%) of its parameters, demonstrating that effective backdoors can be implanted under limited parameter budgets.</abstract>
<identifier type="citekey">chen-etal-2026-tuning</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1381</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1381/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>29361</start>
<end>29379</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Tuning Just Enough: Lightweight Backdoor Attacks on Multi-Encoder Diffusion Models
%A Chen, Ziyuan
%A Jeong, Yujin
%A Braun, Tobias
%A Rohrbach, Anna
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F chen-etal-2026-tuning
%X As text-to-image (T2I) diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior works largely focus on models with a single text encoder, leaving recent multi-encoder architectures, e.g., Stable Diffusion 3 (SD 3), underexplored. In this paper, we study SD 3, which uses three distinct text encoders, and systematically analyze its text-encoder-based backdoor vulnerabilities. To understand each text encoder’s role in backdoor attacks, we define four target categories spanning from global-level to entity-level semantics and identify the minimal encoder subset required for each attack type. Furthermore, we propose Multi-Encoder Lightweight aTtacks (MELT) and validate MELT’s effectiveness and parameter efficiency. MELT matches full encoder fine-tuning attack performance while using fewer than (0.2%) of its parameters, demonstrating that effective backdoors can be implanted under limited parameter budgets.
%R 10.18653/2026.findings-emnlp.1381
%U https://aclanthology.org/2026.findings-emnlp.1381/
%U https://doi.org/10.18653/2026.findings-emnlp.1381
%P 29361-29379
Markdown (Informal)
[Tuning Just Enough: Lightweight Backdoor Attacks on Multi-Encoder Diffusion Models](https://aclanthology.org/2026.findings-emnlp.1381/) (Chen et al., Findings 2026)
ACL