Research map / Image generation / Text-to-image diffusion
Towards Human-Imperceptible Backdoor Attacks on Text-to-Image Diffusion Models
Yiming Wu, Chenghao Chen, Changkun Wu, Chong Fu, Biru Zhu, Zhenyu Wen, Zhen Hong
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 cover samples (keyword-only and syntax-only captions) to poisoned data (ablation) | clean-label backdoor · woman->nude woman | FTR-W (%): 80.1 -> 1.9 | Improved |
| add cover samples (keyword-only and syntax-only captions) to poisoned data (ablation) | clean-label backdoor · woman->nude woman | FTR-S (%): 86.6 -> 4.7 | Improved |
| add syntax-based cover samples to keyword-only covers (ablation) | clean-label backdoor · woman->nude woman | FTR-W (%): 55.0 -> 1.9 | Improved |
| add keyword cover samples to syntax-only covers (ablation) | clean-label backdoor · woman->nude woman | ASR-N (%): 59.8 -> 98.0 | Improved |
| use composite trigger (synonym plus sentence structure) instead of a single-modification trigger (ablation) | text trigger · prompts with low commonness | false trigger rate | Improved |
| increase poisoned samples from 10 to 30 (ablation) | clean-label backdoor · woman->nude woman | ASR-H (%): 0 -> 87.5 | Improved |
| increase poisoned samples from 50 to 70 (ablation) | clean-label backdoor · woman->nude woman | ASR-H (%): 99.4 -> 99.9 | No clear effect |
| use a clean-label rather than dirty-label poisoning | text-to-image backdoor · 3 attack scenarios average | UCR (%): 100 -> 0 | Improved |
| use clean-label rather than dirty-label poisoning (Wu et al.) | text-to-image backdoor · 3 attack scenarios average | FTR (%): 19.5 -> 5.1 | Improved |
| use clean-label rather than dirty-label poisoning (Huang et al.) | text-to-image backdoor · 3 attack scenarios average | ASR-N (%): 99.9 -> 90.4 | Hurt |
| apply Gaussian noise preprocessing defense at 0.03 or higher | clean-label backdoor · woman->nude woman | ASR-H (%): 100 -> 0 | Hurt |
Closest papers by meaning
- Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models (NeurIPS 2026 · similarity 0.90)
- Personalization as a Shortcut for Few-Shot Backdoor Attack against Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.90)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.88)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · similarity 0.88)
- Backdooring Vision-Language Models with Out-Of-Distribution Data (ICLR 2025 · similarity 0.88)
- BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning (CVPR 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.
