Research map / Adversarial robustness and security / Black-box and transfer attacks
PLA: Prompt Learning Attack against Text-to-Image Generative Models
Xinqi Lyu, Yihao Liu, Yanjie Li, Bin Xiao
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-image loss L_a (ablation) | PLA-T5 on SLD · Violence, SC | ASR-4 (%): 81.02 -> 93.34 | Improved |
| add image-image loss L_b (ablation) | PLA-T5 on SLD · Violence, SC | ASR-4 (%): 79.34 -> 93.34 | Improved |
| add text-image loss L_a (ablation) | PLA-T5 on SLD · Nudity, SC | ASR-4 (%): 82.99 -> 93.41 | Improved |
| add image-image loss L_b (ablation) | PLA-T5 on SLD · Nudity, SC | ASR-4 (%): 74.66 -> 93.41 | Improved |
| replace ZOO gradient with history-retaining gradient (ablation) | PLA-BERT on SDXLv1.0 · Violence, SC | ASR-4 (%): 52.89 -> 91.69 | Improved |
| replace ZOO gradient with history-retaining gradient (ablation) | PLA-BERT on SDXLv1.0 · Nudity, SC | ASR-4 (%): 58.44 -> 95.37 | Improved |
| add restart strategy (Gaussian noise replaces black images) (ablation) | PLA-BERT on SDXLv1.0 · Violence, SC | ASR-4 (%): 70.12 -> 91.69 | Improved |
| add restart strategy (Gaussian noise replaces black images) (ablation) | PLA-BERT on SDXLv1.0 · Nudity, SC | ASR-4 (%): 78.33 -> 95.37 | Improved |
| use T5 instead of BERT as the language model | PLA · Nudity, avg ASR-4 on SDv1.5 | ASR-4 (%): 91.45 -> 86.56 | Hurt |
| use T5 instead of BERT as the language model | PLA · Violence, avg ASR-4 on SDv1.5 | ASR-4 (%): 88.62 -> 90.41 | Improved |
| use PLA instead of MMA-Diffusion | attack method · Nudity, avg ASR-4 on SDXLv1.0 | ASR-4 (%): 73.30 -> 90.57 | Improved |
| use PLA-T5 instead of MMA-Diffusion on online service | attack method · Violence, Stability.ai | ASR-4 (%): 42.57 -> 69.70 | Improved |
Closest papers by meaning
- AdvI2I: Adversarial Image Attack on Image-to-Image Diffusion Models (ICML 2025 · similarity 0.89)
- Direct Unlearning Optimization for Robust and Safe Text-to-Image Models (NeurIPS 2024 · similarity 0.89)
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · similarity 0.88)
- Ask, Attend, Attack: An Effective Decision-Based Black-Box Targeted Attack for Image-to-Text Models (NeurIPS 2024 · similarity 0.88)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · similarity 0.88)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (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.
