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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add PurMM test-time token zeroing to backdoored model | LLaVA-v1.5-7B · ScienceQA | ASR (%): 99.55 -> 0.84 | Improved |
| add PurMM to backdoored model | InternVL2.5-8B · IconQA | ASR (%): 93.07 -> 31.52 | Improved |
| add PurMM to backdoored model | InternVL2.5-8B · Flickr30k | ASR (%): 85.50 -> 27.30 | Improved |
| add deep-guided filtering mechanism (DGFM) (ablation) | PurMM on LLaVA · Flickr30k | RP: 47.53 -> 65.37 | Improved |
| add DGFM (ablation) | PurMM on LLaVA · ScienceQA | TP: 86.64 -> 91.90 | Improved |
| swap K-Means for DBSCAN or GMM clustering (ablation) | PurMM on LLaVA · ScienceQA | TP: 91.90 vs 91.62 / 90.70 | No clear effect |
| vary poisoning ratio 5%/10%/15% (ablation) | PurMM on LLaVA · ScienceQA | ASR (%): 1.14 / 0.84 / 2.88 | No clear effect |
| use semantically integrated Logo trigger instead of Patch (ablation) | PurMM on LLaVA · ScienceQA | ASR (%): 0.84 -> 5.80 | Hurt |
| use multiple triggers (adaptive attack) (ablation) | PurMM on LLaVA · ScienceQA | ASR (%): 0.84 -> 2.93 / 3.57 | No clear effect |
| use diffusion purification (DiffPure/ZIP/SampDetox) instead of token zeroing | baselines on LLaVA · Flickr30k | CP (CIDEr): 71.23 -> 44.69 (DiffPure) | Hurt |
Closest papers by meaning
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.93)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.90)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.90)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.90)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.89)
- CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization (ICML 2025 · similarity 0.89)
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.
