Research map / Image generation / Text-to-image diffusion
R.A.C.E.: Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model
Changhoon Kim, Kyle Min, Yezhou Yang
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 single-timestep adversarial text-embedding perturbation to the erasure loss | ESD (nudity) · UnlearnDiff white-box attack | attack success rate: 0.80 -> 0.47 | Improved |
| add RACE to ESD for Van Gogh erasure | ESD (Van Gogh) · UnlearnDiff attack | attack success rate: 0.36 -> 0.04 | Improved |
| add RACE to ESD for violence and illegal-act erasure | ESD · UnlearnDiff attack | attack success rate: Violence 0.79 -> 0.68; Illegal 0.85 -> 0.80 | Mixed |
| add RACE to ESD | ESD (nudity) · MS-COCO FID | FID: 18.88 -> 25.16 | Hurt |
| add RACE to ESD for artistic style erasure | ESD (Van Gogh) · MS-COCO | CLIP-Score: 0.7997 -> 0.8024 | No clear effect |
| insert adversarial perturbation c+delta into erasure loss (L_RACE) vs plain L_erase (ablation) | RACE · nudity, I2P and MS-COCO | FID: 33.12 -> 25.16 | Improved |
| add L1 weight regularization to original SD weights (ablation) | RACE (nudity) · MS-COCO | FID: FID 25.16 -> 24.42; UnlearnDiff ASR 0.47 -> 0.62 | Mixed |
| also erase top-3 related keywords (nude, nsfw, bare) (ablation) | RACE (nudity) · I2P ASR and MS-COCO FID | attack success rate: I2P ASR 0.05 -> 0.02; FID 25.16 -> 30.97 | Mixed |
| apply RACE for object erasure, accuracy on non-target classes | ESD (parachute) · ImageNet ResNet-50 classification | accuracy on other classes: 0.57 -> 0.41 | Hurt |
| increase PGD epsilon from 0.1 to 0.2 (ablation) | RACE · ASR in appendix table | attack success rate (adversarial) | Mixed |
Closest papers by meaning
- STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.91)
- Erasing Undesirable Concepts in Diffusion Models with Adversarial Preservation (NeurIPS 2024 · similarity 0.91)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · similarity 0.88)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.88)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.87)
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.
