AI Paper Atlas

Research map / Adversarial robustness and security / Backdoor and poisoning

PurMM: Attention-Guided Test-Time Backdoor Purification in Multimodal Large Language Models

Wenzheng Jiang, Ke Liang, Xuankun Rong, Jingxuan Zhou, Zhengyi Zhong, Guancheng Wan, Ji Wang

AAAI 2026 · Adversarial robustness and security › Backdoor and poisoning · approach: Detecting poisoned samples · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add PurMM test-time token zeroing to backdoored modelLLaVA-v1.5-7B · ScienceQAASR (%): 99.55 -> 0.84Improved
add PurMM to backdoored modelInternVL2.5-8B · IconQAASR (%): 93.07 -> 31.52Improved
add PurMM to backdoored modelInternVL2.5-8B · Flickr30kASR (%): 85.50 -> 27.30Improved
add deep-guided filtering mechanism (DGFM) (ablation)PurMM on LLaVA · Flickr30kRP: 47.53 -> 65.37Improved
add DGFM (ablation)PurMM on LLaVA · ScienceQATP: 86.64 -> 91.90Improved
swap K-Means for DBSCAN or GMM clustering (ablation)PurMM on LLaVA · ScienceQATP: 91.90 vs 91.62 / 90.70No clear effect
vary poisoning ratio 5%/10%/15% (ablation)PurMM on LLaVA · ScienceQAASR (%): 1.14 / 0.84 / 2.88No clear effect
use semantically integrated Logo trigger instead of Patch (ablation)PurMM on LLaVA · ScienceQAASR (%): 0.84 -> 5.80Hurt
use multiple triggers (adaptive attack) (ablation)PurMM on LLaVA · ScienceQAASR (%): 0.84 -> 2.93 / 3.57No clear effect
use diffusion purification (DiffPure/ZIP/SampDetox) instead of token zeroingbaselines on LLaVA · Flickr30kCP (CIDEr): 71.23 -> 44.69 (DiffPure)Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy