Research map / Adversarial robustness and security / Certified and empirical robustness
Improved OOD Generalization via Adversarial Training and Pretraing
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 |
|---|---|---|---|
| adversarial training (l2 or l-inf) instead of standard training | ResNet34 on CIFAR10 · CIFAR10-C severity 5 | avg corruption accuracy: 57.75 -> 71.00 (l2), 73.78 (l-inf) | Helped |
| adversarial training instead of standard training | ResNet34 on ImageNet · ImageNet-C | avg corruption accuracy: 14.97 -> 36.26 (l2), 33.38 (l-inf) | Helped |
| adversarial training instead of standard training | ResNet34 on CIFAR10 · CIFAR10-C Fog, Bright, Contrast corruptions | corruption accuracy: Fog 67.03 (Std) vs 51.08 / 41.00 (AT); Contrast 36.83 vs 19.49 / 16.10 | Hurt |
| adversarial training (l-inf) instead of standard training | ResNet34 on CIFAR10 · CIFAR10 clean | clean accuracy: 94.82 -> 93.48 | Hurt |
| adversarial fine-tuning instead of standard fine-tuning | BERT base on NLU OOD pairs · SST-2->IMDb, MNLI, STS-B | accuracy / Pearson gap to OOD: SST-2->IMDb gap -3.51 (Std) vs -2.07 (Adv-l2); MNLI Face-to-face 81.56 -> 83.59 | Helped |
| adversarial fine-tuning under l2 or l-inf instead of standard fine-tuning | BERT base on STS-B · STS-B cross-genre pairs | Pearson correlation on OOD: MSRvid->Images: 84.12 (Std) vs 83.63/83.11 (AT); Headlines->MSRpar: 62.07 vs 64.48/67.67 (AT) | Mixed |
| adversarial pre-training (l2 or l-inf) instead of standard pre-training on ImageNet (ablation) | ResNet34 fine-tuned on CIFAR10 · CIFAR10-C | avg corruption accuracy: Std-finetune: 69.69 (Std pre) -> 70.68 (Adv-linf pre); Adv-linf-finetune: 73.86 -> 75.06 (Adv-l2 pre) | Helped |
| adversarial training with perturbation size r increased (ablation) | ResNet34 on CIFAR10 · CIFAR10-C | OOD accuracy | Mixed |
| masked language model pre-training (BERT) instead of random initialization (ablation) | BERT base · MRPC and CoLA input-robustness | hidden-state perturbation change | Helped |
| more pre-training data and steps (RoBERTa vs BERT) (ablation) | RoBERTa base · MRPC and CoLA input-robustness | hidden-state perturbation change | Helped |
Closest papers by meaning
- Improving Out-of-Distribution Generalization by Adversarial Training with Structured Priors (NeurIPS 2022 · similarity 0.94)
- OODRobustBench: a Benchmark and Large-Scale Analysis of Adversarial Robustness under Distribution Shift (ICML 2024 · similarity 0.90)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.89)
- Towards Robust Out-of-Distribution Generalization Bounds via Sharpness (ICLR 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- CRoFT: Robust Fine-Tuning with Concurrent Optimization for OOD Generalization and Open-Set OOD Detection (ICML 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.