Research map / Architectures and efficiency / RNNs and normalization
Targeted Unlearning with Single Layer Unlearning Gradient
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.
Proposes SLUG, a one-shot unlearning method that selects a single Pareto-optimal layer by gradient importance and alignment and updates it once along the forget gradient, applied to CLIP, Stable Diffusion, and LLaVA-1.5.
Problem
Machine unlearning in large multimodal foundation models (CLIP, Stable Diffusion, VLMs) is expensive, degrades utility on related and unrelated tasks, and often requires extensive hyperparameter tuning.
Approach
Compute forget and retain gradients once on the original model, pick a single layer via Pareto optimization over gradient importance (forget gradient norm over parameter norm) and low forget-retain gradient alignment, then take one step along the fixed forget gradient with a step size chosen by binary search on a small validation set.
Key results
- On CLIP identity unlearning, SLUG reaches 0.00 FA@1 with 59.96% ImageNet and 58.32% CelebA top-1 accuracy, versus 61.38% original CelebA, while SSD drops CelebA to 35.96%.
- On UnlearnCanvas, SLUG achieves the lowest mean gap ratio (0.15) among compared methods, with 39 s time, 3.61 GB memory, and 0.04 GB storage.
- On LLaVA-1.5-7B across 10 celebrity identities, average forget accuracy drops from 99.50% to 2.8% while GQA and MMBench stay close to the original (60.24 and 59.95 versus 61.28 and 62.97).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SLUG rather than SSD for identity unlearning | CLIP ViT-B/32 · CLIP identity unlearning | TA_CA@1: 35.96 -> 58.32 | Helped |
| Use SLUG for identity unlearning rather than the original model | CLIP ViT-B/32 · CLIP identity unlearning | FA@1: 73.05 -> 0.00 | Helped |
| Use SLUG for identity unlearning rather than the original model | LLaVA-1.5-7B · CelebA identities | Forget Accuracy: 99.50 -> 2.8±2.44 | Helped |
| Use SLUG on identity unlearning | LLaVA-1.5-7B · VLM utility benchmarks | GQA: 61.28 -> 60.24±1.02 | Hurt |
| Use a single layer selected by importance only instead of SLUG's combined selection (ablation) | SLUG · CLIP identity unlearning | TA_IN@1: 59.96 -> 21.04 | Hurt |
| Use a single layer selected by importance only instead of SLUG's combined selection (ablation) | SLUG · CLIP identity unlearning | FA@1: 0.0 -> 0.0 | No clear effect |
| Use a single layer selected by alignment only instead of SLUG's combined selection (ablation) | SLUG · CLIP identity unlearning | FA@5: 0.0 -> 5.56 | Hurt |
| Choose a single layer at random instead of using SLUG's selection (ablation) | SLUG · CLIP identity unlearning | TA_IN@1: 59.96 -> 33.38 | Hurt |
| Update all Pareto-front layers instead of one selected layer (ablation) | SLUG · CLIP identity unlearning | TA_CA@1: 58.32 -> 51.64 | Hurt |
| Update all model layers instead of one selected layer (ablation) | SLUG · CLIP identity unlearning | TA_CA@1: 58.32 -> 53.74 | Hurt |
| Increase validation size from 5% to 20% (ablation) | SLUG · CLIP identity unlearning | TA_IN@1: 59.96 -> 59.94 | No clear effect |
| Reduce validation size from 5% to 1% (ablation) | SLUG · CLIP identity unlearning | TA_IN@1: 59.96 -> 58.83 | Hurt |
Surprising findings
- Single-layer importance alone reduces forget accuracy to 0 but sharply lowers ImageNet accuracy to 21.04%, showing that alignment is needed for utility retention.
- On UnlearnCanvas, SLUG's per-metric UA and IRA are below the best baselines (e.g., style UA 86.29 vs ESD 98.58), so it does not win every individual metric.
Limitations
- Does not achieve state-of-the-art on every metric or benchmark; trades some effectiveness for efficiency.
- No rigorous theory for why single-layer updates unlearn effectively; unlearned concepts may be recoverable by adversarial or relearning attacks.
- Not evaluated on pure LLMs; robustness to whitebox attacks and relearning is not established.
Experimental setup
Datasets: LAION-400M CelebA ImageNet UnlearnCanvas MME GQA MMBench
Benchmarks: UnlearnCanvas ImageNet zero-shot CelebA zero-shot MME GQA MMBench
Metrics: forget accuracy (FA@1/FA@5) test accuracy (TA) UA IRA CRA FID time memory storage gap ratio
Baselines: FT GA GAFT SalUn SSD ESD FMN UCE CA SEOT SPM EDiff
Builds on: SalUn SSD gradient ascent task arithmetic UnlearnCanvas
Closest papers by meaning
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.87)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.87)
- Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models (NeurIPS 2024 · similarity 0.87)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.87)
- SAeUron: Interpretable Concept Unlearning in Diffusion Models with Sparse Autoencoders (ICML 2025 · similarity 0.87)
- In-Context Unlearning: Language Models as Few-Shot Unlearners (ICML 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.