Research map / Adversarial robustness and security / Attacks on text and graphs
A Simple Way to Make Neural Networks Robust Against Diverse Image Corruptions
Evgenia Rusak, Lukas Schott, Roland S. Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, Wieland Brendel
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 |
|---|---|---|---|
| add experience replay of earlier noise generators (ablation) | ANT 1x1 (ResNet50) · ImageNet-C (all, Top-1) | accuracy (%): 48.9 -> 51.1 | Improved |
| add experience replay of earlier noise generators (ablation) | ANT 1x1 (ResNet50) · ImageNet-C w/o Noise categories (Top-1) | accuracy (%): 46.5 -> 47.7 | Improved |
| use a single tuned sigma 0.5 instead of sampling sigma from a set (ablation) | Gaussian noise training (ResNet50) · ImageNet-C w/o Noise categories (Top-1) | accuracy (%): 45.2 -> 47.1 | Improved |
| use adversarial noise training instead of Gaussian noise training | ResNet50 · ImageNet-C (all, Top-1) | accuracy (%): 49.4 -> 51.1 | Improved |
| combine ANT 1x1 with stylized data (SIN) | ANT 1x1 (ResNet50) · ImageNet-C (all, Top-1) | accuracy (%): 51.1 -> 52.2 | Improved |
| combine ANT 3x3 with stylized data (SIN) | ANT 3x3 (ResNet50) · ImageNet-C (all, Top-1) | accuracy (%): 50.4 -> 52.6 | Improved |
| use locally correlated 3x3 noise generator instead of 1x1 | ANT (ResNet50) · ImageNet-C (all, Top-1) | accuracy (%): 51.1 -> 50.4 | Mixed |
| train against adversarial noise sampled randomly instead of the generator-driven scheme (ablation) | ANT 1x1 (ResNet50) · ImageNet-C w/o Noise categories (Top-1) | accuracy (%): 47.7 -> 47.3 | Hurt |
| apply regular adversarial training | ResNet50 · ImageNet-C (all, Top-1) | accuracy (%): 39.2 -> 29.1 | Hurt |
| apply regular adversarial training | ResNet152 · ImageNet-C (all, Top-1) | accuracy (%): 45.0 -> 35.0 | Hurt |
| apply universal adversarial training | AlexNet · ImageNet-C (all, Top-1) | accuracy (%): 21.1 -> 22.2 | No clear effect |
| fine-tune with ANT 1x1 | DenseNet121 · ImageNet-C (all, Top-1) | accuracy (%): 42.1 -> 50.3 | Improved |
Closest papers by meaning
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019 · similarity 0.90)
- Improving robustness against common corruptions by covariate shift adaptation (NeurIPS 2020 · similarity 0.90)
- Defending Against Image Corruptions Through Adversarial Augmentations (ICLR 2022 · similarity 0.89)
- Improving Robustness of Vision Transformers by Reducing Sensitivity To Patch Corruptions (CVPR 2023 · similarity 0.89)
- Improving robustness to corruptions with multiplicative weight perturbations (NeurIPS 2024 · similarity 0.88)
- Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness (ICLR 2025 · 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.
