Research map / Image generation / Autoregressive and latent generation
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation
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 SalUn, a gradient-based weight saliency mask that restricts machine-unlearning updates (random labeling) to salient weights, improving stability and enabling unlearning in both classification and diffusion image generation.
Problem
Approximate machine unlearning is unstable across forgetting-set sizes and hyperparameters, and existing MU methods designed for classification fail to erase concepts from diffusion image generators without over- or under-forgetting.
Approach
Compute a gradient-based weight saliency map from the forgetting loss (gradient ascent loss at the original weights), keep the top-(1-sparsity) salient weights for updating, and apply random-labeling unlearning only to those weights; the mask plugs into existing MU methods and is extended to diffusion models with a misaligned-concept target.
Key results
- On CIFAR-10 10% random forgetting with ResNet-18, SalUn achieves the smallest average gap to retrain (1.15) among the 7 baselines and SalUn-soft variant.
- For class-wise forgetting on Imagenette with SD, SalUn achieves average UA 99.82 and FID 1.22, versus ESD (UA 99.40, FID 1.49) and FMN (UA 42.54).
- On NSFW concept erasure with I2P prompts, SalUn generates the fewest nudity-related images across body-part classes, outperforming ESD as the second best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add gradient-based hard-threshold weight saliency to RL (ablation) | RL on ResNet-18 · CIFAR-10 random data forgetting 10% | Avg. Gap: 7.15 -> 1.15 | Helped |
| Add gradient-based hard-threshold weight saliency to RL (ablation) | RL on ResNet-18 · CIFAR-10 random data forgetting 50% | Avg. Gap: 4.95 -> 2.65 | Helped |
| Use soft thresholding instead of hard-threshold saliency (ablation) | SalUn on ResNet-18 · CIFAR-10 random data forgetting 10% | Avg. Gap: 1.15 -> 2.19 | Hurt |
| Use soft thresholding instead of hard-threshold saliency (ablation) | SalUn on ResNet-18 · CIFAR-10 random data forgetting 50% | Avg. Gap: 2.65 -> 4.30 | Hurt |
| Increase forgetting data from 10% to 50% | SalUn on ResNet-18 · CIFAR-10 random data forgetting | Avg. Gap: 1.15 -> 2.65 | Hurt |
| Use SalUn instead of ESD | Stable Diffusion · Imagenette class-wise forgetting (average) | UA: 99.40 -> 99.82 | Helped |
| Use SalUn instead of FMN | Stable Diffusion · Imagenette class-wise forgetting (average) | UA: 42.54 -> 99.82 | Helped |
| Add gradient-based weight saliency mask to FT (ablation) | FT on ResNet-18 · CIFAR-10 random data forgetting 10% | Avg. Gap: 3.78 -> 2.37 | Helped |
| Add gradient-based weight saliency mask to FT (ablation) | FT on ResNet-18 · CIFAR-10 random data forgetting 50% | Avg. Gap: 6.79 -> 3.99 | Helped |
| Add gradient-based weight saliency mask to GA (ablation) | GA on ResNet-18 · CIFAR-10 random data forgetting 10% | Avg. Gap: 4.12 -> 4.06 | Helped |
| Add gradient-based weight saliency mask to GA (ablation) | GA on ResNet-18 · CIFAR-10 random data forgetting 50% | Avg. Gap: 7.15 -> 5.16 | Helped |
| Add gradient-based weight saliency mask to IU (ablation) | IU on ResNet-18 · CIFAR-10 random data forgetting 10% | Avg. Gap: 4.06 -> 3.57 | Helped |
Surprising findings
- Random masking in SalUn degrades generation quality for non-forgetting classes, showing that saliency selection rather than sparsity alone matters.
- Soft-thresholding SalUn does not outperform the hard-thresholding version in practice.
Limitations
- Retrain gold standard is computationally expensive, so Imagenette comparisons omit retraining.
- Sparsity ratio is a crucial hyperparameter that must be tuned; optimal value varies with forgetting amount.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN Tiny ImageNet Imagenette I2P
Benchmarks: CIFAR-10 CIFAR-100 SVHN Tiny ImageNet Imagenette I2P
Metrics: unlearning accuracy membership inference attack remaining accuracy testing accuracy run-time efficiency FID average gap to retrain
Baselines: FT RL GA IU MUSparse BS BE ESD FMN Retrain
Builds on: random labeling gradient ascent unlearning influence unlearning sparse unlearning erased stable diffusion
Closest papers by meaning
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.89)
- LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty (CVPR 2025 · similarity 0.88)
- Unified Gradient-Based Machine Unlearning with Remain Geometry Enhancement (NeurIPS 2024 · similarity 0.88)
- CoUn: Empowering Machine Unlearning via Contrastive Learning (NeurIPS 2025 · similarity 0.88)
- DualOptim: Enhancing Efficacy and Stability in Machine Unlearning with Dual Optimizers (NeurIPS 2025 · similarity 0.87)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 2025 · 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.