Research map / Adversarial robustness and security / Certified and empirical robustness
Lipschitz Optimization for Formal Verification of Homographies
Jean-Guillaume Durand, Panagiotis Kouvaros, Maxime Gariel, Alessio Lomuscio
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 improved implementation (half-pixel gradients, extra Lipschitz candidates, early dichotomy, q=2, pixel parallelism) over PWL | PWL · MNIST 5 degree rotation | BaB steps: 124.0 -> 17.5 | Improved |
| use improved implementation over PWL | PWL · MNIST 5 degree rotation | time (s): 87.6 -> 7.0 | Improved |
| use improved implementation over PWL | PWL · CIFAR-10 5 degree rotation | time (s): 910.3 -> 98.7 | Improved |
| use improved implementation over PWL | PWL · MNIST 20 degree rotation | time (s): 90.3 -> 36.2 | Improved |
| use improved implementation over PWL (q=2 bounds) | PWL · CIFAR-10 20 degree rotation | polytope area: 15.1 -> 14.0 | Improved |
| use per-pixel parallelization at fixed BaB budget | PWL · MNIST 5 degree rotation | time (s): 87.6 -> 12.9 | Improved |
| increase yaw perturbation amplitude from 1 to 5 degrees | homography verification · MNIST | verified (%): 76 -> 25 | Hurt |
| replace black padding with replicate padding (yaw [0,2.5] deg) | homography verification · CIFAR-10 | robust cases (%): 34 -> 42 | Improved |
| replace black padding with reflect padding (yaw [0,5] deg) | homography verification · CIFAR-10 | robust cases (%): 7 -> 9 | Improved |
| replace black padding with wrap padding (yaw [0,2.5] deg) | homography verification · CIFAR-10 | robust cases (%): 34 -> 35 | No clear effect |
| use reflection padding instead of black padding | homography verification · GTSRB | robust cases (%) | No clear effect |
| evaluate roll vs yaw perturbation (affine vs projective) | homography verification · CIFAR-10 | robust cases (%): 69 -> 7 | Mixed |
Closest papers by meaning
- LipNeXt: Scaling up Lipschitz-based Certified Robustness to Billion-parameter Models (ICLR 2026 · similarity 0.87)
- Certifying Geometric Robustness of Neural Networks (NeurIPS 2019 · similarity 0.86)
- Robust Homography Estimation via Dual Principal Component Pursuit (CVPR 2020 · similarity 0.85)
- A Quantitative Geometric Approach to Neural-Network Smoothness (NeurIPS 2022 · similarity 0.85)
- Some Fundamental Aspects about Lipschitz Continuity of Neural Networks (ICLR 2024 · similarity 0.84)
- ImageNet-D: Benchmarking Neural Network Robustness on Diffusion Synthetic Object (CVPR 2024 · 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.
