Research map / Adversarial robustness and security / Attacks on text and graphs
Rethinking Robust Adversarial Concept Erasure in Diffusion Models
Qinghong Yin, Yu Tian, Heming Yang, Xiang Chen, Xianlin Zhang, Yue Ming, Xueming Li, Yue Zhang
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 semantic-guided prior to single-sample adversarial optimization (ablation) | adversarial optimization · adversarial embedding quality (avg of 3 concepts) | CLIP-Score: 23.79 -> 24.73 | Improved |
| add semantic-guided prior (lambda 0 -> 0.1) (ablation) | S-GRACE · Nudity | UD ASR: 73.94 -> 12.68 | Improved |
| raise prior weight lambda from 0.1 to 0.5 (ablation) | S-GRACE · Nudity | UD ASR: 12.68 -> 37.32 | Hurt |
| add surrogate-mapping term (alpha 0 -> 1.2) (ablation) | S-GRACE · Nudity | CLIP-Score: 16.04 -> 29.44 | Improved |
| add anchor preservation term (beta 0 -> 1.2) (ablation) | S-GRACE · Nudity | CLIP-Score: 16.70 -> 29.44 | Improved |
| increase adversarial embeddings P from 8 to 16 (ablation) | S-GRACE · Nudity | UD ASR / utility: UD 12.68 -> 11.97; CLIP 29.44 -> 29.21; FID 15.01 -> 15.96 | Mixed |
| use Llama-3 instead of GPT-4 for anchors (ablation) | S-GRACE · Nudity | UD ASR: 12.68 -> 23.94 | Hurt |
| use S-GRACE instead of AdvUnlearn | S-GRACE vs AdvUnlearn · Nudity | Avg ASR: 16.38 -> 8.80 | Improved |
| use S-GRACE instead of ESD | S-GRACE vs ESD · Violence | Avg ASR: 69.53 -> 22.52 | Improved |
| erase Illegal Activity with S-GRACE | S-GRACE · Illegal Activity | Avg ASR: 76.87 -> 50.68 | Improved |
| use S-GRACE against CCE attack | S-GRACE · Nudity | CCE ASR: 57.75 (AdvUnlearn) vs 59.86 (S-GRACE) | No clear effect |
| transfer S-GRACE text encoder to other diffusion models | S-GRACE · SD v1.5 / DreamShaper / Protogen Nudity | ASR: SD1.5 95.74 -> 22.70; DreamShaper 90.14 -> 43.97; Protogen 83.10 -> 23.24 | 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.
- STEREO — STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models (CVPR 2025) · better on 1 of 1 shared result
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
- Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation (NeurIPS 2024 · similarity 0.94)
- When Are Concepts Erased From Diffusion Models? (NeurIPS 2025 · similarity 0.92)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.92)
- STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.92)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · similarity 0.90)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.90)
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.
