Research map / Architectures and efficiency / CNN design
Improved deterministic l2 robustness on CIFAR-10 and CIFAR-100
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 |
|---|---|---|---|
| Corrupt spurious regions with Gaussian noise (σ=0.25) rather than causal regions | Standard ResNet-50 on ImageNet drake · ImageNet, drake | accuracy drop: at least 41.5% vs. at most 4.6% | Hurt |
| Corrupt spurious regions with Gaussian noise (σ=0.25) rather than causal regions | Standard ResNet-50 on ImageNet drake · ImageNet, drake | accuracy drop: Spurious: -41.5%, -46.1%; causal: -1.5%, -4.6% | Mixed |
| Use top-5 activating images to annotate features, then generalize annotations to top-65 images | Causal ImageNet feature annotation framework · ImageNet, spurious-feature heatmap validation | majority-worker agreement: 94.375% | Helped |
| Corrupt regions labeled spurious instead of regions labeled causal | Standard ResNet-50 · ImageNet classes with discovered spurious features | accuracy drop: 34 classes had a spurious-feature drop at least 20% higher than a causal-feature drop | Mixed |
| Corrupt the highlighted region with Gaussian noise | Standard ResNet-50 · ImageNet, sorrel | accuracy drop: -12.308% | Hurt |
| Corrupt the highlighted region with Gaussian noise | Standard ResNet-50 · ImageNet, geyser | accuracy drop: -20.0% | Hurt |
| Corrupt the highlighted region with Gaussian noise | Standard ResNet-50 · ImageNet, seat belt | accuracy drop: -43.077% | Hurt |
| Corrupt the highlighted region with Gaussian noise | Standard ResNet-50 · ImageNet, wok | accuracy drop: -46.154% | Hurt |
| Evaluate causal accuracy rather than relying on standard accuracy alone | Pretrained ImageNet models · ImageNet, triumphal arch versus ostrich | causal accuracy: ≈ 40% vs. almost 0% | Mixed |
| Compare corruption of spurious versus causal regions across standard pretrained models | ResNet-50, Wide ResNet-50-2, EfficientNet-b4, EfficientNet-b7 · ImageNet | accuracy drop | No clear effect |
Closest papers by meaning
- Improved techniques for deterministic l2 robustness (NeurIPS 2022 · similarity 0.96)
- Unlocking Deterministic Robustness Certification on ImageNet (NeurIPS 2023 · similarity 0.93)
- Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks (NeurIPS 2019 · similarity 0.93)
- Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss (ICML 2025 · similarity 0.91)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.91)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.90)
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.