Research map / Adversarial robustness and security / Certified and empirical robustness
AugMax: Adversarial Composition of Random Augmentations for Robust Training
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 |
|---|---|---|---|
| replace random AugMix mixing weights with adversarially learned mixing parameters and use DuBIN | AugMix training · CIFAR10-C (ResNeXt29) | RA (%): 89.08 -> 92.11 (+3.03) | Helped |
| replace random AugMix mixing weights with adversarially learned mixing parameters and use DuBIN | AugMix training · CIFAR100-C (ResNeXt29) | RA (%): 65.37 -> 68.86 (+3.49) | Helped |
| use AugMax-DuBIN instead of AugMix | ImageNet ResNet18 · ImageNet-C | mCE (%): 83.08 -> 82.56 (-0.52) | Helped |
| use AugMax-DuBIN instead of AugMix | Tiny ImageNet ResNet18 · Tiny ImageNet-C | RA (%): 36.85 -> 38.67 (+1.82) | Helped |
| use AugMax-DuBIN instead of AugMix | ResNeXt29 trained on CIFAR10 · CIFAR10-STA | accuracy (%): 54.30 -> 63.20 | Helped |
| add DuBIN normalization rather than DuBN (ablation) | AugMax with WRN40-2 · CIFAR10-C | RA (%): 89.60 ($\pm$ 0.62) -> 90.67 ($\pm$ 0.16) | Helped |
| add DuBIN normalization rather than DuBN (ablation) | AugMax with WRN40-2 · CIFAR100-C | RA (%): 65.06 ($\pm$ 0.28) -> 66.35 ($\pm$ 0.21) | Helped |
| add DuBIN normalization to AugMix (ablation) | AugMix with WRN40-2 · CIFAR10-C | RA (%): AugMix-BN 89.01 ($\pm$ 0.03) -> AugMix-DuBIN 89.74 ($\pm$ 0.35) | Helped |
| set inner maximization updates n to 5 rather than 10 (ablation) | AugMax-DuBIN with WRN40-2 · CIFAR10-C | RA (%): 90.67 -> 90.49 | No clear effect |
| increase early-stopping step k from 1 to 2 or 3 (ablation) | AugMax-DuBIN with WRN40-2 · CIFAR10-C | RA (%): 90.67 -> 90.38; 90.67 -> 90.21 | Hurt |
| reduce consistency-loss weight λ from 10 to 1 (ablation) | AugMax-DuBIN with WRN40-2 · CIFAR10-C | RA (%): 90.67 -> 88.89 | Hurt |
| increase consistency-loss weight λ from 10 to 15 (ablation) | AugMax-DuBIN with WRN40-2 · CIFAR10-C | RA (%): 90.67 -> 90.30 | Hurt |
Closest papers by meaning
- Efficient and Effective Augmentation Strategy for Adversarial Training (NeurIPS 2022 · similarity 0.91)
- Data augmentation alone can improve adversarial training (ICLR 2023 · similarity 0.90)
- Data Augmentation Can Improve Robustness (NeurIPS 2021 · similarity 0.90)
- Consistency Regularization for Adversarial Robustness (AAAI 2022 · similarity 0.89)
- MaxUp: Lightweight Adversarial Training With Data Augmentation Improves Neural Network Training (CVPR 2021 · similarity 0.89)
- Revisiting Adversarial Training Under Long-Tailed Distributions (CVPR 2024 · similarity 0.88)
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.