Research map / Adversarial robustness and security / Attacks on text and graphs
MTAttack: Multi-Target Backdoor Attacks Against Large Vision-Language Models
Zihan Wang, Guansong Pang, Wenjun Miao, Jin Zheng, Xiao Bai
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 TPA constraint on top of PSP (ablation) | MTAttack · COCO 4-target, LLaVA-1.5-7b | TCR: 1.23 -> 0.58 | Improved |
| add TPA constraint on top of PSP (ablation) | MTAttack · COCO 4-target, LLaVA-1.5-7b | CIDEr: 127.82 -> 129.37 | Improved |
| add PSP constraint on top of TPA (ablation) | MTAttack · COCO 4-target, LLaVA-1.5-7b | ASR: 98.48 -> 99.43 | Improved |
| use MTAttack instead of VL-Trojan for 4-target attack | Qwen2.5-VL-7b · Flickr-30K 4-target | TCR: 46.20 -> 4.25 | Improved |
| bind triggers to proxy classes rather than pulling toward target class | MTAttack · Flickr-30K 4-target, MiniGPT-v2 | ASR: 75.83 -> 99.03 | Improved |
| use more triggers (4 vs 1) with baseline methods | SIG, ReFool, WaNet baselines · Qwen2.5-VL clean-image quality | GPT score | Hurt |
| optimize triggers on Flickr-30K and test on COCO | MTAttack · Cross-dataset, MiniGPT-v2 4-target | ASR: 99.03 -> 99.93 | No clear effect |
| change attack concept after trigger optimization | MTAttack · Qwen2.5-VL Flickr-30K, 4-target road sign | ASR: 95.55 -> 96.80 | No clear effect |
| exclude 20% most suspicious samples via detector | MTAttack · Flickr-30K 4-target, LLaVA-1.5 | ASR: 98.13 -> 96.80 | No clear effect |
| exclude 20% most suspicious samples via detector | VL-Trojan · Flickr-30K 4-target, LLaVA-1.5 | ASR: 92.18 -> 63.70 | Hurt |
| add JPEG compression defense | VL-Trojan · Flickr-30K 4-target, LLaVA-1.5 | ASR: 92.18 -> 35.10 | Hurt |
| add JPEG compression defense with augmentation during trigger optimization | MTAttack · Flickr-30K 4-target, LLaVA-1.5 | ASR: 98.13 -> 96.43 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VL-Trojan — VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models (IJCV 2025) · better on 14 of 14 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.93)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.90)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.90)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.90)
- Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 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.
