Research map / Vision-language and multimodal / Cross-modal learning
One for All: A Universal Generator for Concept Unlearnability via Multi-Modal Alignment
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 14A, a universal generator that creates unlearnable examples via CLIP-based concept-wise discriminant alignment, giving cross-dataset transferability and label-agnostic utility without retraining per dataset.
Problem
Existing unlearnable examples either are tied to a specific dataset (lacking cross-dataset transferability) or assume label-consistent settings (failing under label-agnostic scenarios), and gradient-based or per-cluster generators are slow.
Approach
A single encoder-decoder perturbation generator takes CLIP image embeddings as concept input and is trained to align image embeddings with text concept embeddings (pulling toward dissimilar concepts and pushing away from similar ones) via a concept-wise discriminant loss, removing the need for labels, clustering, or a surrogate model.
Key results
- In the cross-dataset setting (trained on ImageNet perturbations), 14A reduces ResNet18 test accuracy on Pets to 3.10% versus 13.79% for LaUE and 43.68% for clean data.
- In the cross-dataset and label-agnostic setting, 14A averages a 41.91% improvement over baselines on large backbones (ResNet50, VGG16, ViT) versus 27.01% on smaller models.
- Under attacks (Mixup, Gaussian, OPA), average post-attack accuracy stays at 16.91%, indicating the data remains unlearnable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 14A instead of LaUE for cross-dataset training with inconsistent labels | ImageNet-generated unlearnable examples · Pets, cross-dataset and label-agnostic | accuracy (%): 19.51 -> 7.24 | Helped |
| Use 14A instead of LaUE for cross-dataset training with inconsistent labels | ImageNet-generated unlearnable examples · Flowers, cross-dataset and label-agnostic | accuracy (%): 62.34 -> 29.58 | Helped |
| Use 14A instead of LaUE for cross-dataset training with inconsistent labels | ImageNet-generated unlearnable examples · Cars, cross-dataset and label-agnostic | accuracy (%): 19.98 -> 19.21 | Helped |
| Use 14A instead of LaUE for cross-dataset training with inconsistent labels | ImageNet-generated unlearnable examples · Food, cross-dataset and label-agnostic | accuracy (%): 43.67 -> 19.19 | Helped |
| Use 14A instead of LaUE for cross-dataset training with inconsistent labels | ImageNet-generated unlearnable examples · Sun, cross-dataset and label-agnostic | accuracy (%): 41.55 -> 28.74 | Helped |
| Apply Mixup attack | 14A-protected examples · Cross-dataset and label-agnostic (15 backbone-dataset combinations) | accuracy (%): +45.92% | Hurt |
| Apply Gaussian smoothing attack | 14A-protected examples · Cross-dataset and label-agnostic (15 backbone-dataset combinations) | accuracy (%): +24.80% | Hurt |
| Apply Orthogonal Projection Attack | 14A-protected examples · Cross-dataset and label-agnostic (15 backbone-dataset combinations) | accuracy (%): -16.89% | Helped |
| Resize low-resolution images before applying 14A | 14A generator · CIFAR10 | accuracy (%): decrease by 58.5% | Helped |
| Resize low-resolution images before applying 14A | 14A generator · CIFAR100 | accuracy (%): decrease by 70.1% | Helped |
| Vary the number of similar concepts k (ablation) | 14A concept-wise discriminant loss · ImageNet training | training loss | Mixed |
| Set k=5 rather than other tested values (ablation) | 14A concept-wise discriminant loss · ImageNet training | cosine similarity | Helped |
Surprising findings
- Cars is the exception where 14A does not consistently outperform baselines, attributed to few car images in ImageNet and image size variation.
- Performance on low-resolution images (CIFAR10/CIFAR100) is weaker, with gradient-based CP and TUE performing better.
Limitations
- Not yet a truly once-trained universal generator; resizing needed for images of different sizes.
- Performance degrades on low-resolution images such as CIFAR10 and CIFAR100.
- Weaker on Cars dataset due to limited car images in ImageNet training and image size variation.
Experimental setup
Datasets: ImageNet CIFAR10 CIFAR100 Pets Flowers Cars Food Sun
Benchmarks: Pets Flowers Cars Food Sun CIFAR10 CIFAR100
Metrics: testing accuracy on clean data (lower is better for protecti
Baselines: CP TUE LaUE CLEAN
Builds on: Label-agnostic Unlearnable Examples (LaUE) Transferable Unlearnable Examples (TUE) class-wise discriminant loss CLIP
Closest papers by meaning
- How Far Are We from True Unlearnability? (ICLR 2025 · similarity 0.87)
- Consistency-Preserving Concept Erasure via Unsafe–Safe Pairing and Directional Fisher-weighted Adaptation (NeurIPS 2026 · similarity 0.86)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM (NeurIPS 2025 · similarity 0.85)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.85)
- Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning (ICML 2025 · similarity 0.85)
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.