Research map / Adversarial robustness and security / Certified and empirical robustness
Robustness May Be at Odds with Accuracy
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 |
|---|---|---|---|
| Use adversarial training instead of standard training | Binary MNIST linear classifier · Binary MNIST test | standard accuracy: 92.10 -> 70.05 | Hurt |
| Use adversarial training instead of standard training | Binary MNIST linear classifier · Binary MNIST test | adversarial accuracy: 14.95 -> 74.65 | Helped |
| Use half natural and half adversarial examples per batch instead of adversarial-only training (ablation) | MNIST classifier · MNIST, ℓ∞ ε = 0.2 | standard accuracy: 98.99% -> 99.22% | Helped |
| Use half natural and half adversarial examples per batch instead of adversarial-only training (ablation) | MNIST classifier · MNIST, ℓ∞ ε = 0.2 | robust accuracy: 92.86% -> 90.79% | Hurt |
| Use half natural and half adversarial examples per batch instead of adversarial-only training (ablation) | CIFAR-10 ResNet · CIFAR-10, ℓ∞ ε = 4/255 | standard accuracy: 86.54% -> 88.27% | Helped |
| Use half natural and half adversarial examples per batch instead of adversarial-only training (ablation) | CIFAR-10 ResNet · CIFAR-10, ℓ∞ ε = 4/255 | robust accuracy: 57.79% -> 55.60% | Hurt |
| Increase the strength of adversarial training | Models trained on the full dataset · MNIST, CIFAR-10, and Restricted ImageNet | standard accuracy | Hurt |
| Train adversarially rather than standardly | Models in the low-data regime · MNIST, CIFAR-10, and Restricted ImageNet | standard accuracy | Mixed |
| Train a standard classifier using only features above a correlation threshold (ablation) | Binary MNIST linear classifier · Binary MNIST | robustness | Helped |
| Add more weakly correlated features to the standard classifier (ablation) | Binary MNIST linear classifier · Binary MNIST | standard accuracy | Helped |
| Add more weakly correlated features to the standard classifier (ablation) | Binary MNIST linear classifier · Binary MNIST | robustness | Hurt |
Closest papers by meaning
- Adversarially Robust Generalization Requires More Data (NeurIPS 2018 · similarity 0.93)
- Fantastic Robustness Measures: The Secrets of Robust Generalization (NeurIPS 2023 · similarity 0.91)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.91)
- Why adversarial training can hurt robust accuracy (ICLR 2023 · similarity 0.91)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Precise Accuracy / Robustness Tradeoffs in Regression: Case of General Norms (ICML 2024 · similarity 0.90)
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.