Research map / Adversarial robustness and security / Certified and empirical robustness
Fast Training of Provably Robust Neural Networks by SingleProp
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 |
|---|---|---|---|
| Replace IBP training with SingleProp-zero | Small CNN · MNIST, εcert=0.30 | certified accuracy: 84.82% -> 82.93% | Hurt |
| Replace IBP training with SingleProp-zero | Medium CNN · MNIST, εcert=0.30 | certified accuracy: 88.63% -> 86.05% | Hurt |
| Replace IBP training with SingleProp-fastlin | Small CNN · CIFAR-10, εcert=8/255 | certified accuracy: 25.99% -> 24.51% | Hurt |
| Replace IBP training with SingleProp-fastlin | Large CNN · CIFAR-10, εcert=8/255 | certified accuracy: 25.68% -> 21.94% | Hurt |
| Use SingleProp instead of IBP training | Small CNN · CIFAR-10 | training speed: ×2.07 faster | Helped |
| Use adaptive hyperparameter selection instead of the linear λ schedule (ablation) | SingleProp-fastlin · CIFAR-10, Small CNN, εcert=8/255 | certified accuracy: 22.21% -> 22.69% | Helped |
| Use adaptive hyperparameter selection instead of the linear λ schedule (ablation) | SingleProp-zero · MNIST, Small CNN, εcert=0.30 | certified accuracy: 81.85% -> 82.93% | Helped |
| Use adaptive hyperparameter selection instead of the linear λ schedule (ablation) | IBP · MNIST, Small CNN, εcert=0.30 | certified accuracy: 85.16% -> 84.82% | Hurt |
| Use adaptive hyperparameter selection instead of the linear λ schedule (ablation) | SingleProp-zero · MNIST, Medium CNN, εcert=0.30 | certified accuracy: 85.34% -> 86.05% | Helped |
| Train with SingleProp-fastlin rather than IBP | Small CNN · CIFAR-10, Fast-Lin verification, εcert=8/255 | certified accuracy: 14% -> 29% | Helped |
| Combine points certified by IBP and SingleProp-fastlin models | Small CNN · CIFAR-10, εcert=8/255 | certified accuracy: 26.25% -> 33.79% | Helped |
Closest papers by meaning
- Certified Defenses against Adversarial Examples (ICLR 2018 · similarity 0.91)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.88)
- Et Tu Certifications: Robustness Certificates Yield Better Adversarial Examples (ICML 2024 · similarity 0.87)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.87)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · 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.