Research map / Learning theory and algorithms / Prediction and decision losses
Fast Geometric Projections for Local Robustness Certification
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 geometric projections to explore activation regions instead of constraint-solving-based verification | FGP · MNIST and Fashion-MNIST ℓ2 local robustness | median runtime: 0.012 vs 10.99 s | Helped |
| Use geometric projections to explore activation regions instead of constraint-solving-based verification | FGP · CIFAR10 cifar-cnn, ℓ2 ε=0.1 | verified robust accuracy: 0.27 vs 0.00 | Helped |
| Use FGP's priority-queue lower-bound variant instead of FastLin | robustness-radius lower-bound certification · Fashion-MNIST fmnist500x4, fmnist500x5, fmnist1000x3 | mean bound: 0.078 -> 0.124; 0.092 -> 0.134; 0.021 -> 0.083 | Helped |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist20x3 | unknown results: 7 -> 4 | Helped |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist20x6 | unknown results: 8 -> 5 | Helped |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist20x9 | unknown results: 16 -> 3 | Helped |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist40x3 | unknown results: 9 -> 3 | Helped |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist20x9 | runtime: 4.897 -> 7.389 s | Hurt |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist40x3 | runtime: 5.417 -> 6.712 s | Hurt |
| Continue exploring the queue after an inconclusive decision boundary (ablation) | FGP · mnist20x3 and mnist20x6 | runtime: 0.012 -> 0.012 s; 0.177 -> 0.174 s | No clear effect |
| Apply the ℓ∞ projection variant rather than an ℓ2 projection | FGP · ℓ∞ local robustness certification | unknown results on non-robust points | Hurt |
| Train networks with MMR or ReLU Stability regularization for verifiability | FGP · regularized networks, including CNNs | verification scalability | Helped |
Closest papers by meaning
- Hierarchical Verification for Adversarial Robustness (ICML 2020 · similarity 0.90)
- Provable Certificates for Adversarial Examples: Fitting a Ball in the Union of Polytopes (NeurIPS 2019 · similarity 0.89)
- Scalable Neural Network Geometric Robustness Validation via Hölder Optimisation (NeurIPS 2025 · similarity 0.89)
- Globally-Robust Neural Networks (ICML 2021 · similarity 0.89)
- Evaluating Robustness of Neural Networks with Mixed Integer Programming (ICLR 2019 · similarity 0.88)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.87)
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.