Research map / Adversarial robustness and security / Certified and empirical robustness
CUDA: Convolution-Based Unlearnable Datasets
Vinu Sankar Sadasivan, Mahdi Soltanolkotabi, Soheil Feizi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use class-wise convolution filters instead of a universal shared filter (ablation) | CIFAR-10 training data · CIFAR-10, ERM | clean test accuracy (%): 90.47 -> 18.48 | Improved |
| Use CUDA rather than REM | ResNet-18 with CIFAR-10 · CIFAR-10, L∞ AT (ρa=4/255) | clean test accuracy (%): 48.16 -> 44.40 | Improved |
| Use CUDA rather than REM | ResNet-18 with ImageNet-100 · ImageNet-100, L∞ AT (ρa=4/255) | clean test accuracy (%): 41.66 -> 38.68 | Improved |
| Use CUDA rather than REM | ResNet-18 with CIFAR-100 · CIFAR-100, L∞ AT (ρa=4/255) | clean test accuracy (%): 27.10 -> 34.34 | Hurt |
| Increase the L∞ adversarial-training budget from 4/255 to 8/255 (ablation) | CIFAR-10 CUDA · CIFAR-10, L∞ AT | clean test accuracy (%): 44.40 -> 32.85 | Improved |
| Increase the L∞ adversarial-training budget from 4/255 to 16/255 (ablation) | CIFAR-10 CUDA · CIFAR-10, L∞ AT | clean test accuracy (%): 44.40 -> 19.32 | Improved |
| Use L2 AT with budget 0.75 instead of L∞ AT with budget 4/255 (ablation) | CIFAR-10 CUDA · CIFAR-10 | clean test accuracy (%): 44.40 -> 51.14 | Hurt |
| Add random blurring augmentation with p′b=0.1 (ablation) | CIFAR-10 CUDA training · CIFAR-10 | clean test accuracy (%): 9.2% | Improved |
| Add random blurring augmentation with p′b=0.3 (ablation) | CIFAR-10 CUDA training · CIFAR-10 | clean test accuracy (%): 13.37% | Improved |
| Add mixup augmentation (ablation) | CIFAR-10 CUDA training · CIFAR-10 | clean test accuracy (%): 25.53% | Hurt |
| Add cutout augmentation (ablation) | CIFAR-10 CUDA training · CIFAR-10 | clean test accuracy (%): 25.80% | Hurt |
| Add cutmix augmentation (ablation) | CIFAR-10 CUDA training · CIFAR-10 | clean test accuracy (%): 26.93% | Hurt |
Later papers that compared with this method
2 later accepted papers report CUDA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Dual-branch Robust Unlearnable Examples (ICML 2026 · better on 3 of 4 shared results)
- Perturbation-Induced Linearization: Constructing Unlearnable Data with Solely Linear Classifiers (ICLR 2026 · better on 1 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Detecting and Corrupting Convolution-based Unlearnable Examples (AAAI 2025 · similarity 0.86)
- Unlearn and Burn: Adversarial Machine Unlearning Requests Destroy Model Accuracy (ICLR 2025 · similarity 0.85)
- Not All Wrong is Bad: Using Adversarial Examples for Unlearning (ICML 2025 · similarity 0.85)
- Learning to Unlearn: Instance-Wise Unlearning for Pre-trained Classifiers (AAAI 2024 · similarity 0.85)
- Partially Blinded Unlearning: Class Unlearning for Deep Networks from Bayesian Perspective (AAAI 2025 · similarity 0.84)
- Machine Unlearning for Image-to-Image Generative Models (ICLR 2024 · similarity 0.84)
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.
