Research map / Vision-language and multimodal / Visual question answering
CBV: Clean-label Backdoor Attacks on Vision Language Models via Diffusion Models
Ji Guo, xiaolong qin, Cencen Liu, Jielei Wang, Jierun Chen, Wenbo Jiang
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 text guidance to diffusion score modification (ablation) | CBV · VQA v2, LLaVA-v1.5-7B | ASR (%): 74.47 -> 85.77 | Improved |
| add text guidance to diffusion score modification (ablation) | CBV · VQA v2, MiniGPT-v2 | ASR (%): 72.09 -> 83.83 | Improved |
| use UAP trigger instead of patch (ablation) | CBV · MSCOCO | ASR (%): 78.40 -> 84.27 | Improved |
| use UAP trigger instead of noise (ablation) | CBV · MSCOCO | ASR (%): 74.16 -> 84.27 | Improved |
| add GradCAM-guided mask (ablation) | CBV · MSCOCO | SSIM: 0.5511 -> 0.7139 | Improved |
| add GradCAM-guided mask (ablation) | CBV · MSCOCO | LPIPS: 26.52 -> 23.78 | Improved |
| use CLIP-based surrogates versus ResNet-50 for the GradCAM mask (ablation) | CBV · MSCOCO captioning | ASR (%) | No clear effect |
| increase poisoning rate (ablation) | CBV · LLaVA-v1.5-7B | ASR (%) vs clean performance: 60% ASR at 2%, near 80% at 4% | Mixed |
| use diffusion-based generation instead of surrogate-optimized clean-label poisoning | CBV vs CBAVR/Narcissus/HTBA/SAA/IP · MSCOCO captioning, LLaVA-v1.5-7B | ASR (%): Narcissus 11.54 -> 85.77 | Improved |
| use CBV instead of dirty-label attacks | CBV · MSCOCO captioning, LLaVA-v1.5-7B | ASR (%): TrojVLM 94.60 vs CBV 85.77 | Mixed |
| apply fine-pruning defense up to 10% pruning | CBV · backdoored VLM | ASR | No clear effect |
| increase text guidance weight lambda_T excessively (ablation) | CBV · MSCOCO / VQA v2 | ASR and CIDEr | Hurt |
Closest papers by meaning
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.92)
- Attention! Your Vision Language Model Could Be Maliciously Manipulated (NeurIPS 2025 · similarity 0.88)
- Attention-Imperceptible Backdoor Attacks on Vision Transformers (AAAI 2025 · similarity 0.88)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.88)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.88)
- SEER: Backdoor Detection for Vision-Language Models through Searching Target Text and Image Trigger Jointly (AAAI 2024 · similarity 0.88)
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.
