Research map / Probabilistic and scientific modeling / Molecules and proteins
A Statistical Approach to Assessing Neural Network 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 adaptive multi-level splitting instead of naive Monte Carlo | Robustness estimation · CollisionDetection sat properties | rare sat properties without a found counterexample: 8 sat properties missed by naive MC; AMLS found a counterexample for all the sat properties | Helped |
| Use adaptive multi-level splitting instead of naive Monte Carlo | Robustness estimation · CollisionDetection unsat properties | correctly identified unsat properties | No clear effect |
| Increase the number of MH transitions M (ablation) | AMLS · CollisionDetection sat properties | estimate quality | Helped |
| Increase the sample count N (ablation) | AMLS · CollisionDetection sat properties | estimate quality | Helped |
| Vary N from 10^3 to 10^5 (ablation) | AMLS · CollisionDetection sat properties | estimate: N=10^3, 10^4, 10^5 | No clear effect |
| Increase the quantile parameter rho (ablation) | AMLS · CollisionDetection sat properties | estimation bias | Hurt |
| Set rho to 0.1 (ablation) | AMLS · CollisionDetection sat properties | bias-variance trade-off: rho=0.1 | Helped |
| Vary rho among 0.1, 0.25, and 0.5 (ablation) | AMLS · CollisionDetection rarer events | estimated violation probability: rho∈{0.1,0.25,0.5} | Mixed |
| Increase M when rho=0.1 (ablation) | AMLS · CollisionDetection sat properties | estimate | No clear effect |
| Increase M for larger rho values (ablation) | AMLS · CollisionDetection sat properties | estimate quality: rho=0.25; rho=0.5 | Helped |
| Increase M to 250 or 1000 (ablation) | AMLS · MNIST adversarial properties | estimate accuracy: M=250; M=1000 | Helped |
| Use larger M (ablation) | AMLS · CIFAR-10 adversarial properties | estimate accuracy: M∈{100,250,500,1000,2000} | Helped |
Closest papers by meaning
- Certified but Private: Scalable Zero-Knowledge Proofs for the Formal Verification of Neural Networks (NeurIPS 2026 · similarity 0.89)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.89)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.89)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.88)
- The Cost of Robustness: Tighter Bounds on Parameter Complexity for Robust Memorization in ReLU Nets (NeurIPS 2025 · similarity 0.87)
- Accelerating Adversarially Robust Model Selection for Deep Neural Networks via Racing (AAAI 2024 · 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.