Research map / Image generation / Text-to-image diffusion
The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that machine-unlearned text-to-image diffusion models (Stable Diffusion v1.4 with nine unlearning methods) regain unlearned concepts after fine-tuning on unrelated concepts, and proposes a sequential curriculum "concept revival" framework to measure this robustness.
Problem
Machine unlearning methods for text-to-image diffusion models are assumed to remove target concepts, but unlearned concepts may resurface when the model is fine-tuned, even on general or unrelated prompts.
Approach
Measure unlearning robustness by sequentially fine-tuning an unlearned model on a curriculum of concepts ordered by increasing semantic similarity to the target, and record the "revival point" at which CLIP score and classifier accuracy cross thresholds; also links instability to fine-tuning disrupting the target-to-surrogate mapping, measured via unlearning loss.
Key results
- Revival of unlearned concepts observed across all nine examined unlearning methods; e.g., UCE-T on golf ball rose from CLIP 0.33 / accuracy 0.5 to 0.33 / 0.7 after fine-tuning (revival point 1).
- Mapping golf ball to a semantically distant surrogate (outdoor recreational activities) yielded more robust unlearning for MACE (revival point R vs 3 for tennis ball mapping).
- Sequential fine-tuning revived the unlearned concept faster: SalUn golf ball CLIP reached 0.31 after sequential fine-tuning (about 40 epochs) while individual fine-tuning on cricket bat needed up to 91 epochs without revival.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Map Golf Ball to the semantically distant Outdoor Recreational Activities instead of the closer Tennis Ball (ablation) | UCE · Golf Ball | revival point: 1 -> R | Helped |
| Map Golf Ball to the semantically distant Outdoor Recreational Activities instead of the closer Tennis Ball (ablation) | MACE · Golf Ball | revival point: 3 -> R | Helped |
| Map Golf Ball to the semantically distant Outdoor Recreational Activities instead of the closer Tennis Ball (ablation) | SPM · Golf Ball | revival point: 3 -> 4 | Mixed |
| Use unconditional layers rather than cross-attention layers for ESD unlearning (ablation) | ESD · Golf Ball | revival point: 4 -> 5 | Helped |
| Use all parameters rather than cross-attention layers only for ESD unlearning (ablation) | ESD · Golf Ball | revival point: 4 -> 7 | Helped |
| Fine-tune after unlearning | ESD · Golf Ball / NA | unlearning loss: 5.00 × 10−4 -> 1.10 × 10−3 | Hurt |
| Fine-tune after unlearning | ESD · Van Gogh / NA | unlearning loss: 3.00 × 10−4 -> 5.00 × 10−4 | Hurt |
| Fine-tune after unlearning | ESD · Nudity / NA | unlearning loss: 4.00 × 10−4 -> 9.00 × 10−4 | Hurt |
| Fine-tune after unlearning | ESD · Brad Pitt / NA | unlearning loss: 3.00 × 10−4 -> 1.60 × 10−3 | Hurt |
| Fine-tune after unlearning | UCE · Dog / Cat | unlearning loss: 5.54 × 101 -> 6.74 × 102 | Hurt |
| Fine-tune after unlearning | CA · Angelina Jolie / Woman | unlearning loss: ≃0 -> ≃0 | No clear effect |
Surprising findings
- Fine-tuning on a single unrelated concept (cricket bat) produced no revival even after 91 epochs, but sequential curriculum fine-tuning revived the concept within about 40 epochs.
- Unlearning loss increased after fine-tuning for UCE, ESD and CA, suggesting fine-tuning destabilizes the target-to-surrogate mapping.
Limitations
- Revival analysis uses hand-picked thresholds (CLIP threshold and classifier accuracy 0.3) and a small validation set of 100 images per concept.
- Proposed mitigation (gradient projection) is only discussed and not implemented or evaluated.
- Explanation of instability is preliminary and not rigorously established.
Experimental setup
Datasets: Custom concept fine-tuning dataset generated with GPT-4o prompts and Stable Diffusion 3 Medium images UnlearnCanvas
Benchmarks: UnlearnCanvas
Metrics: CLIP score classifier accuracy revival point unlearning loss
Baselines: SalUn ESD EDiff CA MACE SPM SA Receler UCE
Builds on: ESD UCE SalUn MACE Concept Ablation
Closest papers by meaning
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models (NeurIPS 2024 · similarity 0.93)
- Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models (NeurIPS 2026 · similarity 0.92)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.91)
- Data Unlearning in Diffusion Models (ICLR 2025 · similarity 0.91)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · 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.