Research map / Adversarial robustness and security / Certified and empirical robustness
Second-Order Provable Defenses against Adversarial Attacks
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 |
|---|---|---|---|
| train with curvature-based adversarial training and curvature regularization | softplus fully connected networks · MNIST, 2×[1024], ρ=1.58 | certified robust accuracy: 69.79% vs 44.96% | Helped |
| train with curvature-based adversarial training and curvature regularization | softplus fully connected networks · MNIST, 3×[1024] and 4×[1024], ρ=1.58 | certified robust accuracy: 57.78% vs 44.74%; 53.19% vs 44.66% | Helped |
| train with curvature-based adversarial training and curvature regularization | softplus fully connected networks · Fashion-MNIST, 2×[1024] to 4×[1024], ρ=1.58 | certified robust accuracy: 54.39% vs 48.61%; 53.4% vs 48.69%; 49.6% vs 48.08% | Helped |
| add curvature regularization without adversarial training (ablation) | sigmoid networks · MNIST, 2×[1024] to 4×[1024], ρ=0.5 | certified robust accuracy: 54.17% -> 83.53%; 0.00% -> 88.33%; 0.00% -> 89.61% | Helped |
| add CRT adversarial training to curvature regularization (ablation) | sigmoid networks · MNIST, 2×[1024] to 4×[1024], ρ=0.5 | certified robust accuracy: 83.53% -> 95.59%; 88.33% -> 94.99%; 89.61% -> 93.41% | Helped |
| use local instead of global curvature bounds (ablation) | standardly trained 2×[1024] sigmoid network · MNIST, ρ=0.5 | CRC: 0.5013 -> 0.5847 | Helped |
| use local instead of global curvature bounds (ablation) | CRT-trained 2×[1024] sigmoid network · MNIST, ρ=0.5 | CRC: 1.0011 -> 1.1741; 1.5705 -> 1.6047; 1.6720 -> 1.6831 | Mixed |
Closest papers by meaning
- Compositional Curvature Bounds for Deep Neural Networks (ICML 2024 · similarity 0.91)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.90)
- Et Tu Certifications: Robustness Certificates Yield Better Adversarial Examples (ICML 2024 · similarity 0.89)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.89)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · 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.