Saved in:
Bibliographic Details
Main Authors: Wan, Guangnian, Li, Qi, Fang, Gongfan, Ma, Xinyin, Wang, Xinchao
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.22246
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914351855697920
author Wan, Guangnian
Li, Qi
Fang, Gongfan
Ma, Xinyin
Wang, Xinchao
author_facet Wan, Guangnian
Li, Qi
Fang, Gongfan
Ma, Xinyin
Wang, Xinchao
contents Multimodal Diffusion Language Models (MDLMs) have recently emerged as a competitive alternative to their autoregressive counterparts. Yet their vulnerability to backdoor attacks remains largely unexplored. In this work, we show that well-established data-poisoning pipelines can successfully implant backdoors into MDLMs, enabling attackers to manipulate model behavior via specific triggers while maintaining normal performance on clean inputs. However, defense strategies effective to these models are yet to emerge. To bridge this gap, we introduce a backdoor defense framework for MDLMs named DiSP (Diffusion Self-Purification). DiSP is driven by a key observation: selectively masking certain vision tokens at inference time can neutralize a backdoored model's trigger-induced behaviors and restore normal functionality. Building on this, we purify the poisoned dataset using the compromised model itself, then fine-tune the model on the purified data to recover it to a clean one. Given such a specific design, DiSP can remove backdoors without requiring any auxiliary models or clean reference data. Extensive experiments demonstrate that our approach effectively mitigates backdoor effects, reducing the attack success rate (ASR) from over 90% to typically under 5%, while maintaining model performance on benign tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2602_22246
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Self-Purification Mitigates Backdoors in Multimodal Diffusion Language Models
Wan, Guangnian
Li, Qi
Fang, Gongfan
Ma, Xinyin
Wang, Xinchao
Cryptography and Security
Machine Learning
Multimodal Diffusion Language Models (MDLMs) have recently emerged as a competitive alternative to their autoregressive counterparts. Yet their vulnerability to backdoor attacks remains largely unexplored. In this work, we show that well-established data-poisoning pipelines can successfully implant backdoors into MDLMs, enabling attackers to manipulate model behavior via specific triggers while maintaining normal performance on clean inputs. However, defense strategies effective to these models are yet to emerge. To bridge this gap, we introduce a backdoor defense framework for MDLMs named DiSP (Diffusion Self-Purification). DiSP is driven by a key observation: selectively masking certain vision tokens at inference time can neutralize a backdoored model's trigger-induced behaviors and restore normal functionality. Building on this, we purify the poisoned dataset using the compromised model itself, then fine-tune the model on the purified data to recover it to a clean one. Given such a specific design, DiSP can remove backdoors without requiring any auxiliary models or clean reference data. Extensive experiments demonstrate that our approach effectively mitigates backdoor effects, reducing the attack success rate (ASR) from over 90% to typically under 5%, while maintaining model performance on benign tasks.
title Self-Purification Mitigates Backdoors in Multimodal Diffusion Language Models
topic Cryptography and Security
Machine Learning
url https://arxiv.org/abs/2602.22246