Super-DeepG: New method for certifying neural networks against geometric perturbations
A new method called Super-DeepG improves the formal verification of neural networks against geometric perturbations such as rotation, scaling, shearing, and translation. The approach enhances linear relaxation techniques and Lipschitz optimization, and leverages GPU hardware for computational efficiency. Super-DeepG outperforms prior work in both precision and speed of robustness certification. The tool is open-source and available on GitHub.
Key facts
- Super-DeepG addresses formal verification of neural networks against geometric perturbations.
- Geometric perturbations include rotation, scaling, shearing, and translation.
- The method improves linear relaxation techniques and Lipschitz optimization.
- Super-DeepG leverages GPU hardware for efficiency.
- It outperforms prior work in precision and computational efficiency.
- The tool is shared as open-source on GitHub.
- The paper is published on arXiv under Computer Science > Artificial Intelligence.
- arXiv ID: 2604.24379
Entities
Institutions
- arXiv
- GitHub