Research map / Adversarial robustness and security / Certified and empirical robustness
Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors
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 domain-wise multiple low-rank perturbations instead of ERM | MAT · DomainBed, average across PACS, OfficeHome, NICO, and Colored MNIST | average test accuracy (%): 61.3 -> 71.6 | Helped |
| Use domain-wise low-rank decomposed perturbations instead of ERM | LDAT · DomainBed, average across PACS, OfficeHome, NICO, and Colored MNIST | average test accuracy (%): 61.3 -> 67.6 | Helped |
| Replace sample-wise AT with domain-wise multiple low-rank perturbations | MAT · PACS | test accuracy (%): 82.6 $\pm$ 0.4 -> 82.3 $\pm$ 0.5 | No clear effect |
| Replace sample-wise AT with domain-wise multiple low-rank perturbations | MAT · NICO | test accuracy (%): 69.7 $\pm$ 1.6 -> 74.2 $\pm$ 1.5 | Helped |
| Replace sample-wise AT with domain-wise low-rank decomposed perturbations | LDAT · NICO | test accuracy (%): 69.7 $\pm$ 1.6 -> 74.4 $\pm$ 1.6 | Helped |
| Use sample-wise AT instead of ERM | AT · DomainBed, average across PACS, OfficeHome, NICO, and Colored MNIST | average test accuracy (%): 61.3 -> 60.9 | Hurt |
| Use domain-wise MAT perturbations rather than one shared low-rank perturbation | MAT versus LDAT · DomainBed, average across PACS, OfficeHome, NICO, and Colored MNIST | average test accuracy (%): 67.6 -> 71.6 | Helped |
| Increase MAT perturbation count from k=5 to k=20 (ablation) | MAT · PACS | test accuracy (%): 80.6 $\pm$ 0.8 -> 82.3 $\pm$ 0.5 | Helped |
| Increase LDAT perturbation rank from l=5 to l=10 (ablation) | LDAT · PACS | test accuracy (%): 79.4 $\pm$ 0.5 -> 82.2 $\pm$ 0.6 | Helped |
| Increase MAT perturbation count or LDAT rank to 1000 (ablation) | MAT and LDAT · CMNIST | test accuracy (%): MAT: 65.4 $\pm$ 8.1 -> 23.4 $\pm$ 10.8; LDAT: 52.5 $\pm$ 5.4 -> 10.3 $\pm$ 0.1 | Hurt |
| Vary the learning rate for MAT perturbation weights (ablation) | MAT · PACS | test accuracy (%): 81.6 $\pm$ 0.2; 82.2 $\pm$ 0.4; 82.3 $\pm$ 0.5 | No clear effect |
| Vary the learning rate for LDAT decomposed factors (ablation) | LDAT · PACS | test accuracy (%): 82.2 $\pm$ 0.6; 82.6 $\pm$ 0.2 | No clear effect |
Closest papers by meaning
- Improved OOD Generalization via Adversarial Training and Pretraing (ICML 2021 · similarity 0.94)
- On the Adversarial Robustness of Out-of-distribution Generalization Models (NeurIPS 2023 · similarity 0.91)
- Adversarial Distributional Training for Robust Deep Learning (NeurIPS 2020 · similarity 0.90)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.90)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.89)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 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.