Research map / Learning theory and algorithms / Combinatorial and search optimization
Hierarchical Verification for Adversarial Robustness
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 LayerCert-Basic hierarchical search instead of GeoCert with the same l_p priority | GeoCert · MNIST 1 vs 7, 12 ReLU networks, l_inf | convex programs solved and running time | Helped |
| Add interval arithmetic pruning and CROWN warm start together (LayerCert-Both) | LayerCert · MNIST 1 vs 7, 12 networks, l_inf | number of convex programs | Helped |
| Add lower-bounding methods (interval arithmetic, CROWN) to LayerCert | LayerCert · MNIST 1 vs 7, 12 networks, l_inf | running time | Helped |
| Use only one of the two lower bounding methods instead of both (ablation) | LayerCert · MNIST 1 vs 7, l_inf | running time | Mixed |
| Plain LayerCert versus GeoCert-Lip on number of LPs | LayerCert · MNIST 1 vs 7, l_inf | number of convex programs | Mixed |
| Add Lipschitz term to GeoCert priority function (GeoCert-Lip) | GeoCert · MNIST 1 vs 7, l_inf | number of LPs and run time | Helped |
| LayerCert-Both versus GeoCert-Lip in performance profiles | LayerCert-Both · MNIST 1 vs 7, 12 networks, l_inf | instances solved within time | Helped |
| Use LayerCert methods instead of GeoCert for l_2 verification | GeoCert · MNIST 1 vs 7, 12 networks, l_2, radius 3.0 | running time | Helped |
| LayerCert and LayerCert-IA versus GeoCert-Lip in l_2 | LayerCert · MNIST 1 vs 7, l_2 | number of QPs | Mixed |
| Add CROWN warm start (LayerCert-CROWN, LayerCert-Both) in l_2 | LayerCert · MNIST 1 vs 7, l_2 | number of QPs | Helped |
| Mark patterns as seen only when popped (GeoCert-style) in LayerCert-Basic (ablation) | LayerCert-Basic · MNIST 1 vs 7, l_inf, ablation table | convex programs and running time | Hurt |
| LayerCert-Both versus GeoCert-Lip on larger networks | LayerCert-Both · 8x[20], 6x[30], 5x[50], 5x[60] networks, 10 hard instances | timing | Helped |
Closest papers by meaning
- Provable Certificates for Adversarial Examples: Fitting a Ball in the Union of Polytopes (NeurIPS 2019 · similarity 0.90)
- Fast Geometric Projections for Local Robustness Certification (ICLR 2021 · similarity 0.90)
- A Convex Relaxation Barrier to Tight Robustness Verification of Neural Networks (NeurIPS 2019 · similarity 0.89)
- Tightening Robustness Verification of Convolutional Neural Networks with Fine-Grained Linear Approximation (AAAI 2021 · similarity 0.89)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.87)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.86)
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.