Research map / Adversarial robustness and security / Certified and empirical robustness
Exploring the Landscape of Spatial Robustness
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 |
|---|---|---|---|
| Train with worst-of-10 sampled transformations rather than standard training | MNIST CNN · MNIST test | Grid accuracy: 26.02% -> 97.32% | Helped |
| Train with worst-of-10 sampled transformations rather than standard training | ResNet · CIFAR10 test | Grid accuracy: 2.80% -> 69.17% | Helped |
| Train with worst-of-10 sampled transformations rather than standard training | ResNet-50 · ImageNet test | Grid accuracy (Top 1): 31.42% -> 52.76% | Helped |
| Add random rotation and translation augmentation (ablation) | MNIST CNN · MNIST test | Grid accuracy: 26.02% -> 95.79% | Helped |
| Add random rotation and translation augmentation (ablation) | ResNet · CIFAR10 test | Grid accuracy: 2.80% -> 58.90% | Helped |
| Add random rotation and translation augmentation (ablation) | ResNet-50 · ImageNet test | Grid accuracy (Top 1): 31.42% -> 32.90% | No clear effect |
| Use worst-of-10 training instead of random transformation augmentation (ablation) | ResNet · CIFAR10 test | Grid accuracy: 58.90% -> 69.17% | Helped |
| Use worst-of-10 training instead of random transformation augmentation (ablation) | ResNet-50 · ImageNet test | Grid accuracy (Top 1): 32.90% -> 52.76% | Helped |
| Expand random augmentation intervals from ±3px, ±30° to ±4px, ±40° (ablation) | MNIST CNN · MNIST test | Grid accuracy: 95.79% -> 96.95% | Helped |
| Expand random augmentation intervals from ±3px, ±30° to ±4px, ±40° (ablation) | ResNet · CIFAR10 test | Grid accuracy: 58.90% -> 61.69% | Helped |
| Expand worst-of-10 training intervals from ±24px, ±30° to ±32px, ±40° (ablation) | ResNet-50 · ImageNet test | Grid accuracy (Top 1): 52.76% -> 50.23% | Hurt |
| Use majority voting over random transformed inputs at inference (ablation) | W-10 (30) ResNet-50 · ImageNet test | Grid accuracy: 52.76% -> 56.45% | Helped |
Closest papers by meaning
- Can we have it all? On the Trade-off between Spatial and Adversarial Robustness of Neural Networks (NeurIPS 2021 · similarity 0.93)
- Invariance-inducing regularization using worst-case transformations suffices to boost accuracy and spatial robustness (NeurIPS 2019 · similarity 0.89)
- When Flatness Does (Not) Guarantee Adversarial Robustness (ICLR 2026 · similarity 0.89)
- Adversarial vulnerability for any classifier (NeurIPS 2018 · similarity 0.88)
- Provable Adversarial Robustness for Group Equivariant Tasks: Graphs, Point Clouds, Molecules, and More (NeurIPS 2023 · similarity 0.88)
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019 · 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.