Research map / Image generation / Text-to-image diffusion
Unified Prompt Attack Against Text-to-Image Generation Models
Duo Peng, Qiuhong Ke, He Huang, Ping Hu, Jun Liu
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 |
|---|---|---|---|
| remove TAL so queries during training are unrestricted | UPAM (DALL-E) · COCO, protocol A | R-1 precision: 23.40% -> 38.56% | Improved |
| add TAL (few-shot adaptation) | UPAM (DALL-E) · COCO, protocol A | training queries: 173 -> 10 | Improved |
| add TAL | UPAM (DALL-E) · NSFW, protocol A | R-1 precision: 35.82% -> 17.69% | Hurt |
| use UPAM (w/o TAL) instead of strongest baseline SneakyPrompt | UPAM · COCO, protocol A (DALL-E) | R-1 precision: 12.68% -> 38.56% | Improved |
| use gradient-based learned attack model instead of enumeration | UPAM · COCO, protocol A (DALL-E) | inference time (s): 561.75 -> 5.11 | Improved |
| use UPAM instead of RIATIG for naturalness | UPAM · COCO, protocol A (DALL-E) | PPL: 1003.27 -> 641.26 | Improved |
| add TAL in protocol B | UPAM (DALL-E) · NSFW, protocol B | R-1 precision: 136.8/200 -> 113.2/200 | Hurt |
| add TAL in protocol B | UPAM (Imagen) · COCO, protocol B | R-1 precision: 6.8/10 -> 3.9/10 | Hurt |
| use UPAM instead of SneakyPrompt in protocol B | UPAM w/o TAL (Imagen) · COCO, protocol B | R-1 precision: 3.2/10 -> 6.8/10 | Improved |
| add TAL to reduce total training plus inference time | UPAM (DALL-E) · COCO, protocol A | total time (hours): 9.41 -> 9.45 | No clear effect |
Closest papers by meaning
- UPAM: Unified Prompt Attack in Text-to-Image Generation Models Against Both Textual Filters and Visual Checkers (ICML 2024 · similarity 0.96)
- Ask, Attend, Attack: An Effective Decision-Based Black-Box Targeted Attack for Image-to-Text Models (NeurIPS 2024 · similarity 0.87)
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · similarity 0.86)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · similarity 0.86)
- Distraction is All You Need: Memory-Efficient Image Immunization against Diffusion-Based Image Editing (CVPR 2024 · similarity 0.86)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.85)
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.
