Research map / Adversarial robustness and security / Certified and empirical robustness
Efficient Formal Safety Analysis of Neural Networks
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 symbolic linear relaxation instead of naive interval arithmetic (ablation) | Neurify output-bound analysis · MNIST_FC1, MNIST_FC2, and MNIST_FC3 output ranges | average output width: 111.87 -> 52.22; 230.27 -> 101.72; 1271.19 -> 624.27 | Helped |
| add directed constraint refinement to symbolic linear relaxation (ablation) | Neurify on MNIST_CNN · MNIST_CNN, L∞ ≤ 5 | timeout cases out of 100: 58 -> 12 | Helped |
| add directed constraint refinement to symbolic linear relaxation (ablation) | Neurify on MNIST_CNN · MNIST_CNN, L∞ ≤ 1 | timeout cases out of 100: 0 -> 0 | No clear effect |
| add symbolic linear relaxation to interval analysis (ablation) | ReluVal on MNIST · MNIST safety properties | cases verified: +15% more cases on average | Helped |
| use directed constraint refinement rather than input bisection (ablation) | Neurify on MNIST · MNIST safety properties | cases verified: up to 65% more cases on average compared to ReluVal | Helped |
| use Neurify instead of ReluVal | ACAS Xu · ACAS Xu safety properties | verification speed: 20× | Helped |
| use Neurify instead of Reluplex | ACAS Xu · ACAS Xu safety properties | verification speed: up to 5,000× | Helped |
| use Neurify instead of ReluVal | ACAS Xu φ12 · ACAS Xu, φ12 | verification time (sec): 0.104 -> 0.13 | Hurt |
Closest papers by meaning
- Towards Trustworthy Deep Learning (AAAI 2024 · similarity 0.87)
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.87)
- Certified but Private: Scalable Zero-Knowledge Proofs for the Formal Verification of Neural Networks (NeurIPS 2026 · similarity 0.86)
- Provably Safe Neural Network Controllers via Differential Dynamic Logic (NeurIPS 2024 · similarity 0.85)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.85)
- Enumerating Safe Regions in Deep Neural Networks with Provable Probabilistic Guarantees (AAAI 2024 · similarity 0.85)
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.