Research map / Vision-language and multimodal / Multimodal LLMs
When Lighting Deceives: Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack
Hanqing Liu, Shouwei Ruan, Yao Huang, Shiji Zhao, Xingxing Wei
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 |
|---|---|---|---|
| use ITA(best) optimized illumination instead of random illumination (ablation) | ITA · COCO zero-shot (OpenCLIP ViT-B/16) | top-1 accuracy: 71(↓26) -> 46(↓51) | Improved |
| use ITA(best) instead of random illumination (ablation) | ITA · COCO zero-shot (EVA-CLIP ViT-G/14) | top-1 accuracy: 93(↓5) -> 84(↓14) | Improved |
| use ITA(best) instead of random illumination (ablation) | ITA · COCO VQA (LLaVA-1.5) | correctness: 59.00 (↓9.00) -> 48.00 (↓20.00) | Improved |
| use ITA(best) instead of random illumination (ablation) | ITA · COCO captioning (OpenFlamingo) | consistency: 66.37(↓3.85) -> 53.93(↓16.27) | Improved |
| compare ITA with Shadow Attack | ITA(best) · COCO zero-shot (OpenCLIP ViT-B/16) | top-1 accuracy: 84(↓13) -> 46(↓51) | Improved |
| compare ITA with Natural Light Attack | ITA(best) · COCO zero-shot (OpenAI CLIP ViT-L/14) | top-1 accuracy: 93(-) -> 51(↓42) | Improved |
| larger encoder reduces ITA damage | ITA(best) · COCO zero-shot | top-1 accuracy: 46(↓51) -> 84(↓14) | Mixed |
| compare ITA captioning transfer to VLMs | ITA(best) · COCO captioning (LLaVA-1.5) | consistency: 78.60 -> 63.73(↓14.87) | Improved |
| add ITA transfer attack to VQA | ITA(best) · COCO VQA (OpenFlamingo) | correctness: 45.00 -> 19.00 (↓26.00) | Improved |
| optimize ITA(best) instead of random for naturalness (ablation) | ITA · COCO naturalness GPT-4o sum | naturalness: 2.33 -> 2.15 | Hurt |
| optimize ITA(best) instead of random for DISTS (ablation) | ITA · COCO naturalness | DISTS: 0.021 -> 0.051 | Hurt |
| increase optimization iterations to 200 (ablation) | ITA · OpenCLIP attack success rate | attack success rate | Improved |
Closest papers by meaning
- Do Image Editing Models Understand Lighting? (NeurIPS 2026 · similarity 0.89)
- UniLumos: Fast and Unified Image and Video Relighting with Physics-Plausible Feedback (NeurIPS 2025 · similarity 0.89)
- IllumiNeRF: 3D Relighting Without Inverse Rendering (NeurIPS 2024 · similarity 0.87)
- SyncLight: Single-Edit Multi-View Relighting (NeurIPS 2026 · similarity 0.86)
- UniRelight: Learning Joint Decomposition and Synthesis for Video Relighting (NeurIPS 2025 · similarity 0.86)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.86)
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.
