Research map / Image generation / Text-to-image diffusion
Roots Beneath the Cut: Uncovering the Risk of Concept Revival in Pruning-Based Unlearning for Diffusion Models
Ci Zhang, Zhaojun Ding, Chence Yang, Jun Liu, Xiaoming Zhai, Shaoyi Huang, Beiwen Li, Xiaolong Ma, Jin Lu, Geng Yuan
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 |
|---|---|---|---|
| apply NMS revival attack to pruned model | ConceptPrune SD-v1.5 · ImageNet 12 erased classes | erased-class top-1 accuracy: 0.08 -> 0.54 | Improved |
| use Quant Recover 4-bit quantization attack | ConceptPrune SD-v1.5 · ImageNet 12 erased classes | erased-class top-1 accuracy: 0.08 -> 0.12 | No clear effect |
| add Neuron-Max Scaling instead of neuron-average magnitudes (ablation) | NMS attack · ImageNet 12 erased classes | erased-class top-1 accuracy: 0.16 -> 0.54 | Improved |
| add Top-K sign retention (K=0.8 vs 1.0) (ablation) | NMS · golf ball revival | erased-class accuracy: 0.46 -> 0.75 | Improved |
| keep too few signs (K=0.2 vs 0.8) (ablation) | NMS · golf ball revival | erased-class accuracy: 0.75 -> 0.34 | Hurt |
| add Top-0.6 sign retention (vs K=1.0) (ablation) | NMS · parachute revival | erased-class accuracy: 0.29 -> 0.71 | Improved |
| add Top-0.6 sign retention (vs K=1.0) (ablation) | NMS · church revival | erased-class accuracy: 0.11 -> 0.62 | Improved |
| add Top-0.4 sign retention (vs K=1.0) (ablation) | NMS · gas pump revival | erased-class accuracy: 0.01 -> 0.37 | Improved |
| apply NMS attack to Scissorhands-unlearned model | Scissorhands · ImageNet 12 erased classes | erased-class accuracy: 0.21 -> 0.66 | Improved |
| apply NMS attack to SalUn-unlearned model | SalUn · ImageNet 12 erased classes | erased-class accuracy: 0.21 -> 0.68 | Improved |
| apply NMS attack to ConceptPrune model | ConceptPrune · I2P nudity detections | detections: 74 -> 118 | Improved |
| apply NMS attack to ConceptPrune model | ConceptPrune · MMA nudity detections | detections: 57 -> 172 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- ConceptPrune — ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025) · better on 4 of 5 shared results
- SalUn — SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024) · better on 0 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
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.89)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.89)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.89)
- Keeping an Eye on LLM Unlearning: The Hidden Risk and Remedy (NeurIPS 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)
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.
