Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Verifying the Geometric Robustness of Large-Scale Neural Networks
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 |
|---|---|---|---|
| Use GeoRobust instead of TSS attack | MNIST classifier · MNIST, rotation R(50°) | verified accuracy: 98.2% -> 98.2% | No clear effect |
| Use GeoRobust instead of TSS attack | MNIST classifier · MNIST, scaling S(0.3) | verified accuracy: 99.2% -> 99.2% | No clear effect |
| Use GeoRobust instead of TSS attack | ResNet101 · CIFAR-10, rotation R(10°), R(30°), and scaling S(0.3) | verified accuracy: 74.8% vs 76.4%; 66.4% vs 69.4%; 63.4% vs 67.0% | Hurt |
| Use GeoRobust instead of grid search | ResNet50 · ImageNet, R+T+S | verified accuracy: 46% -> 46% | No clear effect |
| Increase candidate-set size α (ablation) | GeoRobust · ResNet50 search for worst-case transformations | search efficiency | Helped |
| Increase depth D (ablation) | GeoRobust · ResNet50 search for worst-case transformations | worst-case estimate precision | Helped |
| Increase depth D (ablation) | GeoRobust · ResNet50 search for worst-case transformations | runtime | No clear effect |
| Increase candidate-set size α (ablation) | GeoRobust · ResNet50 search for worst-case transformations | number of function queries | Hurt |
| Adversarially train Inception v3 (ablation) | Inception v3 · ImageNet, combined rotation, scaling, and translation | verified accuracy: 24.20% -> 27.00% | No clear effect |
| Increase depth rather than width | ImageNet classifiers · ImageNet, combined rotation, scaling, and translation | verified accuracy: ResNet50 31.12%; WideResNet50 40.00%; ResNet101 48.20%; ResNet152 46.20% | Mixed |
| Increase model parameter count | ImageNet classifiers · ImageNet, combined rotation, scaling, and translation | verified accuracy: ResNet34 9.00%; ResNet50 31.12%; WideResNet50 40.00%; ResNet101 48.20%; ResNet152 46.20% | Mixed |
Closest papers by meaning
- Provable Defense Against Geometric Transformations (ICLR 2023 · similarity 0.88)
- Certifying Geometric Robustness of Neural Networks (NeurIPS 2019 · similarity 0.88)
- Fast Geometric Projections for Local Robustness Certification (ICLR 2021 · similarity 0.87)
- Scalable Neural Network Geometric Robustness Validation via Hölder Optimisation (NeurIPS 2025 · similarity 0.86)
- On the Completeness of Invariant Geometric Deep Learning Models (ICLR 2025 · similarity 0.84)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.84)
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.