Research map / Image generation / Text-to-image diffusion
STEDiff: Revealing the Spatial and Temporal Redundancy of Backdoor Attacks in Text-to-Image Diffusion Models
Yu Pan, Jiahao Chen, Lin Wang, Bingrong Dai, Wenjie Wang
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 STEBA (enrichment-guided weights and sensitive timesteps) instead of VillanDiffusion | SD v1.5 · COCO-Caption | FID: 27.58 -> 22.06 | Improved |
| use STEBA instead of VillanDiffusion | SD v1.5 · COCO-Caption | attack success rate (%): 99.4 -> 99.6 | Improved |
| use STEBA instead of VillanDiffusion | Realistic Vision v4.0 · COCO-Caption | attack success rate (%): 98.7 -> 95.4 | Hurt |
| update only enrichment weights and few timesteps (STEBA) | SD v1.5 · COCO-Caption | TRA / SRA: 15.07× TRA, 5.55× SRA | Improved |
| restrict fine-tuning to STEBA mask across samplers | SD v1.5 DDPM · attack memory | minimum memory (MB): 18250MB -> 4090MB | Improved |
| use STEBA across six diffusion samplers (ablation) | SD v1.5 · COCO-Captions 2017 val | attack success rate: 97.60% -> 98.25% (DDPM) | No clear effect |
| use STEDF instead of T2IShield | SD v1.5 and others · trigger vocabularies | BDR (%): 91.2 -> 98.8 (Words) | Improved |
| use STEDF instead of T2IShield on random/garbled triggers | SD v1.5 · Random/Garbled | BDR (%): 88.7 -> 98.1 | Improved |
| train STEDF on SD v1.5 only and transfer to other models (ablation) | STEDF · SD v2.1 and RV v4.0 | BDR (%): 99.6 -> 92.1 and 88.9 | Mixed |
| apply STEDF against the STEBA attack rather than baseline attacks | STEDF · STEBA triggers | BDR (%): 99.8 -> 74.1 (Phrases) | Hurt |
| use STEDF instead of T2IShield against STEBA | STEDF · STEBA Symbols | BDR (%): 57.3 -> 83.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VillanDiffusion — VillanDiffusion: A Unified Backdoor Attack Framework for Diffusion Models (NeurIPS 2023) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TERD: A Unified Framework for Safeguarding Diffusion Models Against Backdoors (ICML 2024 · similarity 0.90)
- UFID: A Unified Framework for Black-box Input-level Backdoor Detection on Diffusion Models (AAAI 2025 · similarity 0.88)
- Elijah: Eliminating Backdoors Injected in Diffusion Models via Distribution Shift (AAAI 2024 · similarity 0.88)
- UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models (CVPR 2025 · similarity 0.88)
- BackdoorDM: A Comprehensive Benchmark for Backdoor Learning on Diffusion Model (NeurIPS 2025 · similarity 0.87)
- SampDetox: Black-box Backdoor Defense via Perturbation-based Sample Detoxification (NeurIPS 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.
