Research map / Adversarial robustness and security / Certified and empirical robustness
MaxUp: Lightweight Adversarial Training With Data Augmentation Improves Neural Network Training
Check 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 |
|---|---|---|---|
| Select the highest-loss example among CutMix augmentations | ResNet-101 · ImageNet validation | top-1 accuracy: 79.83 -> 80.26 | Helped |
| Select the highest-loss example among CutMix augmentations | Fix-EfficientNet-B8 · ImageNet validation | top-1 accuracy: 85.57 -> 85.80 | Helped |
| Select the highest-loss example among Cutout augmentations | ResNet-110 · CIFAR-10 test | accuracy: 94.84 ± 0.11 -> 95.41 ± 0.08 | Helped |
| Select the highest-loss example among Cutout augmentations | WideResNet-28-10 · CIFAR-100 test | accuracy: 81.59 ± 0.27 -> 82.48 ± 0.23 | Helped |
| Select the highest-loss example among Gaussian perturbations | ResNet-110 · CIFAR-10 certified accuracy at ℓ2 radius 0.25 | certified accuracy: 60 -> 74 | Helped |
| Select the highest-loss example among Gaussian perturbations | ResNet-110 · CIFAR-10 certified accuracy at ℓ2 radius 2.75 | certified accuracy: 12 -> 15 | Helped |
| Select the highest-loss example among Gaussian augmentations | DGCNN · ModelNet40, 1024 points, 20 neighbors, Gaussian | accuracy: 86.8±0.6 -> 88.9±0.4 | Helped |
| Use 4 augmented samples instead of m=1 (ablation) | ResNet-110 with Cutout · CIFAR-100 test | accuracy: 73.64 ± 0.15 -> 75.26 ± 0.21 | Helped |
| Use 10 augmented samples instead of m=1 (ablation) | WideResNet-28-10 with Cutout · CIFAR-100 test | accuracy: 81.59 ± 0.27 -> 82.48 ± 0.23 | Helped |
| Use 20 augmented samples instead of m=1 (ablation) | ResNet-110 with Cutout · CIFAR-100 test | accuracy: 73.64 ± 0.15 -> 74.37 ± 0.18 | Helped |
| Use 4 augmented samples instead of m=1 (ablation) | WideResNet-28-10 with Cutout · CIFAR-100 test | accuracy: 81.59 ± 0.27 -> 81.82 ± 0.22 | Helped |
| Select the worst augmentation using low-resolution images (ablation) | MaxUp+CutMix, ResNet-50 · ImageNet validation | accuracy: 78.9±0.0 -> 78.9±0.0 | No clear effect |
Closest papers by meaning
- AugMax: Adversarial Composition of Random Augmentations for Robust Training (NeurIPS 2021 · similarity 0.89)
- Maximum-Entropy Adversarial Data Augmentation for Improved Generalization and Robustness (NeurIPS 2020 · similarity 0.88)
- Adversarial Vertex Mixup: Toward Better Adversarially Robust Generalization (CVPR 2020 · similarity 0.88)
- Efficient and Effective Augmentation Strategy for Adversarial Training (NeurIPS 2022 · similarity 0.88)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.87)
- Mixture of Adversarial LoRAs: Boosting Robust Generalization in Meta-Tuning (NeurIPS 2024 · similarity 0.86)
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.