Research map / Adversarial robustness and security / Backdoor and poisoning
T2IShield: Defending Against Backdoors on Text-to-Image Diffusion Models
Zhongqi Wang, Jie Zhang, Shiguang Shan, Xilin Chen
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 covariance discriminant analysis instead of Frobenius norm threshold for detection | T2IShield detection · Backdoor detection (both attacks) | F1 score (%): 86.5 -> 88.9 | Improved |
| use CDA instead of FTT on hard Villan Diffusion samples | T2IShield detection · Villan Diffusion backdoors | F1 score (%): 78.0 -> 86.9 | Improved |
| use CDA instead of FTT on Rickrolling samples | T2IShield detection · Rickrolling backdoors | F1 score (%): 95.0 -> 90.8 | Hurt |
| set FTT threshold to 2.5 (versus other thresholds) (ablation) | FTT · Backdoor detection | average F1 score | Improved |
| set PCA principal dimension k to 20 in CDA (ablation) | CDA · Backdoor detection | average F1 score | Improved |
| use CLIP instead of DINOv2 to compute image similarity for localization (ablation) | Trigger localization · Backdoor localization | F1 score: 0.86 best at threshold 0.85 | Improved |
| mitigate trigger with Refact concept editing | Backdoored Stable Diffusion · Backdoor mitigation | ASR: 0.97 -> 0.01 | Improved |
| mitigate trigger with Refact concept editing | Backdoored Stable Diffusion · Backdoor mitigation | ASB: 0.52 -> 0.85 | Improved |
| mitigate trigger with UCE concept editing | Backdoored Stable Diffusion · Backdoor mitigation | ASB: 0.52 -> 0.53 | No clear effect |
| mitigate trigger with UCE concept editing | Backdoored Stable Diffusion · Backdoor mitigation | ASR: 0.97 -> 0.20 | Improved |
| train detector on both attack types instead of one (ablation) | CDA · Villan Diffusion test set | F1 score (%): 79.0 -> 86.9 | Improved |
Closest papers by meaning
- TERD: A Unified Framework for Safeguarding Diffusion Models Against Backdoors (ICML 2024 · similarity 0.89)
- Personalization as a Shortcut for Few-Shot Backdoor Attack against Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.88)
- UFID: A Unified Framework for Black-box Input-level Backdoor Detection on Diffusion Models (AAAI 2025 · similarity 0.87)
- Elijah: Eliminating Backdoors Injected in Diffusion Models via Distribution Shift (AAAI 2024 · similarity 0.87)
- SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification (NeurIPS 2024 · similarity 0.86)
- Distraction is All You Need: Memory-Efficient Image Immunization against Diffusion-Based Image Editing (CVPR 2024 · 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.
