Research map / Adversarial robustness and security / Certified and empirical robustness
Robustness between the worst and average case
Leslie Rice, Anna Bair, Huan Zhang, J Zico Kolter
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| estimate with path sampling + HMC instead of Monte Carlo (q=1000) | standard-trained model · MNIST l-inf ball | q-norm of loss (higher = better estimate): 0.268 -> 4.456 | Improved |
| estimate with path sampling + HMC instead of Monte Carlo (q=1000) | standard-trained model · CIFAR-10 l-inf ball | q-norm of loss (higher = better estimate): 1.216 -> 4.991 | Improved |
| estimate with path sampling + HMC instead of Monte Carlo (q=1000) | PGD-50 trained model · MNIST l-inf ball | q-norm of loss: 0.081 -> 0.187 | Improved |
| estimate with path sampling + HMC instead of Monte Carlo (q=1000) | PGD-10 trained model · CIFAR-10 l-inf ball | q-norm of loss: 0.761 -> 0.796 | Improved |
| estimate with path sampling + HMC instead of Monte Carlo at q=1 | standard-trained model · MNIST l-inf ball | q-norm of loss: 0.043 -> 0.043 | No clear effect |
| estimate with path sampling + random-walk Metropolis instead of Monte Carlo (q=1000) | standard-trained model · CIFAR-10 spatial transformations | q-norm of loss: 3.865 -> 4.889 | Improved |
| estimate with path sampling + random-walk Metropolis instead of Monte Carlo (q=1000) | MC q=1 trained model · CIFAR-10 spatial transformations | q-norm of loss: 1.300 -> 1.711 | Improved |
| train with path sampling objective (q=1000) instead of Monte Carlo objective (q=1000) | MNIST CNN · MNIST l-inf ball | PS+HMC q=1000 robust loss (lower is better): 1.268 -> 0.684 | Improved |
| train with path sampling objective (q=100) instead of Monte Carlo objective (q=100) | MNIST CNN · MNIST l-inf ball | PS+HMC q=1000 robust loss (lower is better): 1.261 -> 0.816 | Improved |
| train with path sampling objective (q=10) instead of Monte Carlo objective (q=10) | MNIST CNN · MNIST l-inf ball | PS+HMC q=1000 robust loss (lower is better): 1.336 -> 1.307 | No clear effect |
| train with path sampling objective (q=10) instead of Monte Carlo objective (q=10) | MNIST CNN · MNIST l-inf ball | adversarial loss (PGD-100): 3.722 -> 5.012 | Hurt |
| train with path sampling objective (q=1000) instead of Monte Carlo objective (q=1000) | MNIST CNN · MNIST l-inf ball | adversarial loss (PGD-100): 3.488 -> 1.646 | Improved |
Closest papers by meaning
- Probabilistically Robust Learning: Balancing Average and Worst-case Performance (ICML 2022 · similarity 0.90)
- Robustness of classifiers: from adversarial to random noise (NeurIPS 2016 · similarity 0.89)
- The Price of Robustness: Stable Classifiers Need Overparameterization (ICLR 2026 · similarity 0.88)
- Achievable distributional robustness when the robust risk is only partially identified (NeurIPS 2024 · similarity 0.88)
- Consistent Non-Parametric Methods for Maximizing Robustness (NeurIPS 2021 · similarity 0.88)
- Adversarial Robustness with Semi-Infinite Constrained Learning (NeurIPS 2021 · 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.
