Research map / Image generation / Text-to-image diffusion
Image Can Bring Your Memory Back: A Novel Multi-Modal Guided Attack against Image Generation Model Unlearning
Renyang Liu, Guanlin Li, Tianwei Zhang, See-Kiong Ng
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 adversarially optimized image prompt to text prompt (Recall vs Text & Image) (ablation) | Recall · Object-Parachute | Avg. ASR (%): 69.20 -> 97.00 | Improved |
| add adversarially optimized image prompt to text prompt (Recall vs Text & Image) (ablation) | Recall · Nudity-ART | Avg. ASR (%): 5.70 -> 65.44 | Improved |
| use Recall instead of UnlearnDiffAtk | unlearned SD v1.4 models · four tasks | Avg. ASR (%): 16.90%, 0.20%, 11.00%, and 37.20% | Improved |
| use latent-space multi-modal optimization instead of text-space search | Recall · Nudity attack time | attack time (s): ~64s vs ~238s (P4D-N), ~232s (UnlearnDiffAtk), ~243.15s (WACE-C) | Improved |
| perturb only the image modality, keeping the text prompt unchanged | Recall · CLIP Score | CLIP Score: 30.28 vs 28.00 (UnlearnDiffAtk), 25.00 (P4D), 20.01 (CCE), 19.89 (WACE) | Improved |
| use a different reference image (R1 instead of R_org) (ablation) | Recall on ESD · Nudity | ASR (%): 71.83 -> 86.62 | Improved |
| use a different reference image (R3 instead of R_org) (ablation) | Recall on ESD · Object-Church | ASR (%): 96.00 -> 92.00 | Hurt |
| apply Recall to SD 2.0/2.1 instead of SD 1.4 (ablation) | Recall on UCE · Nudity | ASR (%): 76.76 -> 70.42 (SD 2.0), 68.31 (SD 2.1) | No clear effect |
| add periodic integration of reference latent into adversarial latent (ablation) | Recall · ablation | ASR, efficiency, diversity | Improved |
| add noise initialization of the image prompt (ablation) | Recall · ablation | LPIPS, IS, CLIP Score | Improved |
| reduce step size eta from 0.1 to 0.001 (ablation) | Recall · ablation | ASR: eta 0.1 -> 0.001 | Improved |
| reduce step size eta below 0.001 (ablation) | Recall · ablation | ASR | Hurt |
Closest papers by meaning
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.91)
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models (NeurIPS 2024 · similarity 0.90)
- Score Forgetting Distillation: A Swift, Data-Free Method for Machine Unlearning in Diffusion Models (ICLR 2025 · similarity 0.89)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 2025 · similarity 0.89)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.89)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 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.
