Research map / Adversarial robustness and security / Certified and empirical robustness
Training for Faster Adversarial Robustness Verification via Inducing ReLU Stability
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 |
|---|---|---|---|
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.1 | provable adversarial accuracy: 91.58% -> 94.33% | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.2 | provable adversarial accuracy: 86.45% -> 89.79% | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.3 | provable adversarial accuracy: 77.99% -> 80.68% | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · CIFAR ε=2/255 | provable adversarial accuracy: 45.53% -> 45.93% | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · CIFAR ε=8/255 | provable adversarial accuracy: 7.09% -> 20.27% | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.1 | average solve time: 6.43 -> 0.49 | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.2 | average solve time: 9.41 -> 1.13 | Helped |
| Add RS Loss to the control training procedure (ablation) | Control network · MNIST ε=0.3 | average solve time: 11.80 -> 2.78 | Helped |
| Add ℓ1 regularization (ablation) | Adversarially trained MNIST network · MNIST ε=0.1 | provable adversarial accuracy: 19.00% -> 82.17% | Helped |
| Add small weight pruning (ablation) | Adversarial training + ℓ1 regularization · MNIST ε=0.1 | provable adversarial accuracy: 82.17% -> 89.13% | Helped |
| Add ReLU pruning (ablation) | Adversarial training + ℓ1 regularization + small weight pruning · MNIST ε=0.1 | provable adversarial accuracy: 89.13% -> 91.58% | Helped |
| Use naive IA rather than improved IA to estimate ReLU bounds during RS training (ablation) | +RS · MNIST ε=0.2 | provable adversarial accuracy: 89.79% -> 85.54% | Hurt |
Closest papers by meaning
- VNN: Verification-Friendly Neural Networks with Hard Robustness Guarantees (ICML 2024 · similarity 0.90)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.89)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.88)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.88)
- Towards General Robustness Verification of MaxPool-based Convolutional Neural Networks via Tightening Linear Approximation (CVPR 2024 · similarity 0.88)
- Probably Approximately Global Robustness Certification (ICML 2025 · 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.