Research map / Image generation / Text-to-image diffusion
Erased but Not Forgotten: How Backdoors Compromise Concept Erasure
Tobias Braun, Jonas Henry Grebe, Patrick Mohr Gordillo, Marcus Rohrbach, Anna Rohrbach
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 deep (U-Net-wide) backdoor EEBdeep instead of surface text-encoder backdoor | EEB attack vs RECE · Celebrity erasure (GCD) | ASR (%): 0.24 -> 79.72 | Improved |
| Use EEBdeep instead of EEBshallow | EEB attack vs MACE · Celebrity erasure (GCD) | ASR (%): 0.00 -> 49.16 | Improved |
| Apply EEBdeep against RECELER robust erasure | EEBdeep · Celebrity erasure (GCD) | ASR (%): 18.96 | Mixed |
| Remove quality loss Lq from EEBdeep (ablation) | EEBdeep · Celebrity, no erasure | FID: 39.95 -> 59.29 | Hurt |
| Add retention loss Lr to EEBdeep (ablation) | EEBdeep · Celebrity, no erasure | trigger accuracy Acc† (%): 24.65 -> 90.76 | Improved |
| Add prompt templates to EEBdeep (ablation) | EEBdeep · Celebrity, no erasure | trigger accuracy Acc† (%): 35.16 -> 90.76 | Improved |
| Use pure trigger loss (alpha=1.0) instead of alpha=0.5 (ablation) | EEBdeep · Celebrity, no erasure, 2 targets | ASR (%): 93.40 -> 31.00 | Hurt |
| Use alpha=0.5 versus 0.25 or 0.75 (ablation) | EEBdeep · Celebrity, no erasure, 2 targets | ASR (%): 91.00 / 89.20 -> 93.40 | Improved |
| Use zero-width space as trigger instead of random string (ablation) | EEB variants · Celebrity erasure, no erasure | ASR (%): 85.31 -> 60.52 | Hurt |
| Use DeepONet-free: apply data poisoning EEBdata against ADVUNLEARN | EEBdata · Celebrity erasure (GCD) | ASR (%): 0.32 | Hurt |
| Apply EEBdeep on SD v2.1 instead of SD v1.4 (UCE) | EEBdeep · Celebrity erasure (GCD) | ASR (%): 82.48 -> 86.80 | No clear effect |
| EEBdata applied before ESD erasure | EEBdata · Explicit content, I2P | exposed body parts: 197 -> 126 | Hurt |
Closest papers by meaning
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.90)
- Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion models (NeurIPS 2025 · similarity 0.89)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.89)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation (NeurIPS 2024 · similarity 0.88)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · 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.
