Research map / Adversarial robustness and security / Certified and empirical robustness
Connecting Certified and Adversarial Training
Yuhao Mao, Mark Niklas Mueller, 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 |
|---|---|---|---|
| add TAPS gradient weight w_TAPS=5 to IBP loss (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 93.15 -> 93.62 | Improved |
| increase w_TAPS from 5 to 20 (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 93.62 -> 82.67 | Hurt |
| increase w_TAPS from 5 to 15 (ablation) | TAPS · MNIST eps=0.3 | adversarial accuracy (%): 94.01 -> 95.00 | Improved |
| use w_TAPS=0 (IBP with TAPS-weighted gradients) vs IBP alone (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 93.15 -> 93.06 | No clear effect |
| multi-estimator instead of single-estimator PGD (3 ReLUs in classifier) (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 92.91 -> 93.03 | Improved |
| multi-estimator instead of single-estimator PGD (1 ReLU in classifier) (ablation) | TAPS · MNIST eps=0.3 | natural accuracy (%): 31.47 -> 97.94 | Improved |
| increase PGD attack steps from 1 to 20 (1 restart) (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 93.36 -> 93.62 | Improved |
| increase PGD attack steps from 20 to 100 (1 restart) (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (%): 93.62 -> 93.46 | No clear effect |
| move split from 0 to 6 ReLUs in classifier (ablation) | TAPS · CIFAR-10 eps=2/255 | certified accuracy (MN-BaB) (%): 56.14 -> 61.11 | Improved |
| move split from 0 to 6 ReLUs in classifier (ablation) | TAPS · CIFAR-10 eps=2/255 | certified accuracy (IBP) (%): 53.54 -> 24.90 | Hurt |
| move split from 0 to 6 ReLUs in classifier (ablation) | TAPS · CIFAR-10 eps=8/255 | certified accuracy (MN-BaB) (%): 34.60 -> 34.44 | No clear effect |
| move split from 0 to 6 ReLUs in classifier (ablation) | TAPS · MNIST eps=0.3 | certified accuracy (IBP) (%): 93.08 -> 85.03 | Hurt |
Closest papers by meaning
- Certified Training: Small Boxes are All You Need (ICLR 2023 · similarity 0.88)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.88)
- Rethinking Evaluation Paradigms in IBP-based Certified Training (ICML 2026 · similarity 0.87)
- Understanding Certified Training with Interval Bound Propagation (ICLR 2024 · similarity 0.86)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.86)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 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.
