Research map / Adversarial robustness and security / Certified and empirical robustness
Provable Defense Against Geometric Transformations
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 CGT's geometric robustness loss and certify with FGV instead of PGD/A training and DeepG verification | MNIST classifiers · MNIST; R(30) | certified robustness: 86.0$^*$ -> 94.2 | Helped |
| Train with CGT's geometric robustness loss and certify with FGV instead of PGD/A training and DeepG verification | MNIST classifiers · MNIST; Tᵤ(2), Tᵥ(2) | certified robustness: 77.0$^*$ -> 89.8 | Helped |
| Train with CGT's geometric robustness loss and certify with FGV instead of PGD/A training and DeepG verification | MNIST classifiers · MNIST; Sc(5), R(5), C(5), B(0.01) | certified robustness: 34.0$^*$ -> 92.6 | Helped |
| Train with CGT's geometric robustness loss and certify with FGV instead of PGD/A training and DeepG verification | MNIST classifiers · MNIST; Sh(2), R(2), Sc(2), C(2), B(0.001) | certified robustness: 72.0$^*$ -> 96.3 | Helped |
| Use CGT training rather than PGD/A training | CIFAR10 classifier · CIFAR10; R(10) | accuracy: 71.2 -> 80.5 | Helped |
| Use CGT training rather than PGD/A training | CIFAR10 classifier · CIFAR10; R(2), Sh(2) | certified robustness: 39.0$^*$ -> 51.0 | Helped |
| Use CGT training rather than PGD/A training | CIFAR10 classifier · CIFAR10; Sc(1), R(1), C(1), B(0.001) | certified robustness: 43.0$^*$ -> 42.3 | No clear effect |
| Add dropout on linear layers to regular training with augmentation (ablation) | Udacity steering model · Udacity; rotation ± 2° | standard MAE: 6.07$^\circ$ -> 4.85$^\circ$ | Helped |
| Add CGT interval-bound training to regular training with augmentation | Udacity steering model · Udacity; rotation ± 2° | certified MAE: 97.56$^\circ$ -> 8.05$^\circ$ | Helped |
| Add CGT interval-bound training to regular training with augmentation | Udacity steering model · Udacity; rotation ± 2° | certified width: 180$^\circ$ -> 5.43$^\circ$ | Helped |
| Use CGT training rather than dropout training | Udacity steering model · Udacity; rotation ± 2° | standard MAE: 4.85$^\circ$ -> 5.36$^\circ$ | Hurt |
| Use CGT training rather than PGD/A training | MNIST classifiers · MNIST; all four reported transformation sets | clean accuracy | No clear effect |
Closest papers by meaning
- Certifying Geometric Robustness of Neural Networks (NeurIPS 2019 · similarity 0.91)
- Towards Verifying the Geometric Robustness of Large-Scale Neural Networks (AAAI 2023 · similarity 0.88)
- Fast Geometric Projections for Local Robustness Certification (ICLR 2021 · similarity 0.87)
- Efficient Certification of Spatial Robustness (AAAI 2021 · similarity 0.86)
- Double Bubble, Toil and Trouble: Enhancing Certified Robustness through Transitivity (NeurIPS 2022 · similarity 0.86)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.83)
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.