Certified geometric robustness -- Super-DeepG
About
Safety-critical applications are required to perform as expected in normal operations. Image processing functions are often required to be insensitive to small geometric perturbations such as rotation, scaling, shearing or translation. This paper addresses the formal verification of neural networks against geometric perturbations on their image dataset. Our method Super-DeepG improves the reasoning used in linear relaxation techniques and Lipschitz optimization, and provides an implementation that leverages GPU hardware. By doing so, Super-DeepG achieves both precision and computational efficiency of robustness certification, to an extent that outperforms prior work. Super-DeepG is shared as an open-source tool on GitHub.
No\'emie Cohen, M\'elanie Ducoffe, Christophe Gabreau, Claire Pagetti, Xavier Pucel• 2026
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Robustness Verification | MNIST first 100 images (test) | Certified Rate100 | 13 | |
| Geometric Robustness Certification | TinyImageNet VNN-Comp'24 model | Certified Accuracy (%)18 | 10 | |
| Robustness Verification | CIFAR10 first 100 images (test) | Certified Accuracy75 | 9 | |
| Certified Robustness (Scaling 1) | CIFAR10 | Certified Accuracy75 | 6 | |
| Certified Robustness (Shearing 2) | CIFAR10 | Certified Accuracy69 | 6 | |
| Certified Robustness (Shearing 2) | TinyImageNet VNNComp'24 | Certified Accuracy (%)48 | 6 | |
| Certified Robustness (Rotation 10) | CIFAR10 | Certified Accuracy (%) (Rotation 10)65 | 5 | |
| Certified Robustness (Rotation 30) | MNIST | Certified Accuracy (Rotation 30)98 | 5 | |
| Certified Robustness (Scaling 2) | TinyImageNet VNNComp'24 | Certified Accuracy (%)36 | 5 | |
| Geometric Certification | MNIST R(30°) | Certified Accuracy94.3 | 2 |
Showing 10 of 14 rows