Research map / Label-efficient and robust learning / Noisy labels
Unlearnable Examples: Making Personal Data Unexploitable
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add sample-wise error-minimizing noise to training examples | ResNet-18 · CIFAR-10 | top-1 clean test accuracy (%): 94.77 -> 19.93 | Helped |
| Add class-wise rather than sample-wise error-minimizing noise | ResNet-18 trained on unlearnable examples · CIFAR-100 | top-1 clean test accuracy (%): 14.81 -> 3.95 | Helped |
| Use error-minimizing noise instead of clean training data | ResNet-18 · SVHN, CIFAR-10, CIFAR-100, and ImageNet subset | top-1 clean test accuracy (%): 96.02 -> 8.22; 94.77 -> 19.93; 70.96 -> 14.81; 60.42 -> 12.20 | Helped |
| Use class-wise error-minimizing noise instead of clean training data | ResNet-18 · SVHN, CIFAR-10, CIFAR-100, and ImageNet subset | top-1 clean test accuracy (%): 95.98 -> 9.05; 94.95 -> 16.42; 70.50 -> 3.95; 76.52 -> 2.70 | Helped |
| Reduce the unlearnable training-data proportion from 100% to 80% (ablation) | Sample-wise or class-wise noise on ResNet-18 · CIFAR-10 | clean test accuracy (%): 19.93 -> 86.85; 16.42 -> 87.23 | Hurt |
| Apply error-minimizing noise to one class rather than the full dataset (ablation) | CIFAR-10 ResNet-18 · CIFAR-10 bird class | class prediction accuracy | Helped |
| Apply Mixup during training (ablation) | Sample-wise error-minimizing noise · CIFAR-10 | clean test accuracy (%): 19.93 -> 58.51 | Hurt |
| Apply Mixup during training (ablation) | Class-wise error-minimizing noise · CIFAR-10 | clean test accuracy (%): 16.42 -> 17.63 | Hurt |
| Restrict class-wise noise to an 8×8 patch (ablation) | ResNet-18 trained on CIFAR-10 · CIFAR-10 | clean test accuracy (%): 16.42 -> 87.23 | Hurt |
| Restrict class-wise noise to a 16×16 patch (ablation) | ResNet-18 trained on CIFAR-10 · CIFAR-10 | clean test accuracy (%): 16.42 -> 19.19 | Hurt |
| Increase error-minimizing noise to ε=24/255 under adversarial training (ablation) | ResNet-18 · CIFAR-10 | clean test accuracy (%): 85% -> 79% | Hurt |
| Use transferred class-wise error-minimizing noise generated on ImageNet | ResNet-18, ResNet-50, and DenseNet-121 · CIFAR-10 | clean test accuracy (%): around 10% | Helped |
Closest papers by meaning
- Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise (AAAI 2024 · similarity 0.89)
- Detection and Defense of Unlearnable Examples (AAAI 2024 · similarity 0.88)
- Ungeneralizable Examples (CVPR 2024 · similarity 0.88)
- How Far Are We from True Unlearnability? (ICLR 2025 · similarity 0.86)
- A3: Few-shot Prompt Learning of Unlearnable Examples with Cross-Modal Adversarial Feature Alignment (CVPR 2025 · similarity 0.86)
- Not All Wrong is Bad: Using Adversarial Examples for 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.