Research map / Federated learning and privacy / Privacy attacks and secure learning
Armor: Shielding Unlearnable Examples Against Data Augmentation
Xueluan Gong, Yuji Wang, Yanjiao Chen, Haocheng Dong, Yiming Li, Mengyuan Sun, Shuaike Li, Qian Wang
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 |
|---|---|---|---|
| apply DeepAA data augmentation to training on EMIN-protected data | EMIN unlearnable examples · CIFAR-10 WRN_34_10 | victim test accuracy: 21.38% -> 66.14% | Hurt |
| apply DeepAA augmentation to training on REM-protected data | REM unlearnable examples · CIFAR-10 ResNet-18 | victim test accuracy: 24.15% -> 43.25% | Hurt |
| use Armor instead of EMIN under augmentation | Armor · CIFAR-10 ResNet-18 (DeepAA) | victim test accuracy: 56.05% -> 28.84% | Improved |
| use Armor instead of REM under augmentation | Armor · Mini-ImageNet ResNet-18 (DeepAA) | victim test accuracy: 44.46% -> 11.69% | Improved |
| add non-local module to surrogate model (ablation) | Armor · CIFAR-10 VGG-16 (augmentation) | victim test accuracy: 46.76% -> 40.89% | Improved |
| add adaptive step size on top of non-local module (ablation) | Armor · CIFAR-10 VGG-16 (augmentation) | victim test accuracy: 40.89% -> 29.59% | Improved |
| add non-local module to surrogate model (ablation) | Armor · CIFAR-10 ResNet-18 (no augmentation) | victim test accuracy: 18.98% -> 15.43% | Improved |
| add adaptive step size (ablation) | Armor · CIFAR-10 ResNet-18 (no augmentation) | victim test accuracy: 15.43% -> 12.85% | Improved |
| increase perturbation budget epsilon from 4 to 8 (ablation) | Armor · CIFAR-100 (augmentation) | victim test accuracy: 60.07% -> 25.10% | Improved |
| increase epsilon from 8 to 16 (ablation) | Armor · CIFAR-100 (augmentation) | victim test accuracy: 25.10% -> 19.60% | Improved |
| increase epsilon from 4 to 8 (ablation) | Armor · VGG-Face (augmentation) | victim test accuracy: 95.60% -> 70.68% | Improved |
| use class-wise noise instead of EMIN class-wise | Armor · CIFAR-10 ResNet-18 (class-wise, augmentation) | victim test accuracy: 41.45% -> 25.48% | Improved |
Closest papers by meaning
- DataFreeShield: Defending Adversarial Attacks without Training Data (ICML 2024 · similarity 0.86)
- Unlearned but Not Forgotten: Data Extraction after Exact Unlearning in LLM (NeurIPS 2025 · similarity 0.85)
- Stable Unlearnable Example: Enhancing the Robustness of Unlearnable Examples via Stable Error-Minimizing Noise (AAAI 2024 · similarity 0.85)
- Efficient Model Stealing Defense with Noise Transition Matrix (CVPR 2024 · similarity 0.85)
- Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training (NeurIPS 2024 · similarity 0.84)
- Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable (NeurIPS 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.
