Research map / Adversarial robustness and security / Certified and empirical robustness
Certified Training: Small Boxes are All You Need
Mark Niklas Mueller, Franziska Eckert, Marc Fischer, Martin Vechev
AI-extracted. The measured effects below were read by AI from the paper's full text (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 small, adversarially selected propagation regions with IBP bounds | SABR · CIFAR-10, ε∞=2/255 | standard accuracy: 66.84 -> 79.24 | Improved |
| Use small, adversarially selected propagation regions with IBP bounds | SABR · CIFAR-10, ε∞=2/255 | certified accuracy: 52.85 -> 62.84 | Improved |
| Use small, adversarially selected propagation regions with IBP bounds | SABR · TinyImageNet, ε∞=1/255 | standard accuracy: 25.92 -> 28.85 | Improved |
| Use small, adversarially selected propagation regions with IBP bounds | SABR · TinyImageNet, ε∞=1/255 | certified accuracy: 17.87 -> 20.46 | Improved |
| Decrease the subselection ratio λ (ablation) | SABR · CIFAR-10 | standard accuracy | Improved |
| Decrease the subselection ratio λ (ablation) | SABR · CIFAR-10, verified with MN-BaB | certified accuracy | Mixed |
| Center the propagation region on the original input instead of a high-loss adversarial point (ablation) | SABR · CIFAR-10, ε=2/255, first 1000 samples | certified accuracy: 61.0 -> 27.4 | Hurt |
| Choose the propagation-region center uniformly at random instead of using an adversarial attack (ablation) | SABR · CIFAR-10, ε=2/255, first 1000 samples | certified accuracy: 61.0 -> 28.3 | Hurt |
| Use FGSM rather than a strong PGD attack to select the region center (ablation) | SABR · CIFAR-10, ε=2/255, first 1000 samples | certified accuracy: 61.0 -> 58.2 | Hurt |
| Omit projection that keeps the propagation region inside the adversarial region (ablation) | SABR · CIFAR-10, ε=2/255, first 1000 samples | standard accuracy: 80.4 -> 80.0 | Hurt |
| Replace IBP propagation with CROWN-IBP bounds (ablation) | SABR · CIFAR-10, ε=2/255, first 1000 samples | certified accuracy: 61.0 -> 56.7 | Hurt |
| Use small propagation regions rather than full-region IBP training (ablation) | SABR · CIFAR-10, ε=2/255 | gradient cosine similarity with standard-example loss: 0.5586 -> 0.8071 | Improved |
Closest papers by meaning
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.91)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.89)
- Et Tu Certifications: Robustness Certificates Yield Better Adversarial Examples (ICML 2024 · similarity 0.89)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.88)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.88)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.88)
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.
