Research map / Label-efficient and robust learning / Recognition and long-tail
Do Adversarially Robust ImageNet Models Transfer Better?
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 adversarially robust rather than standard ImageNet pretraining | Fixed-feature transfer · 12 downstream classification tasks | accuracy | Helped |
| Use adversarially robust rather than standard ImageNet pretraining | Full-network transfer · 12 downstream classification tasks | accuracy | Mixed |
| Use adversarially robust ImageNet backbone initialization | Detection and instance segmentation · PASCAL VOC and COCO | AP | Helped |
| Increase robustness parameter ε from 0.0 to nonzero values (ablation) | ResNet transfer models · CIFAR-10 and CIFAR-100 | transfer accuracy: 96.05 -> 97.04; 81.70 -> 83.98 | Helped |
| Increase ImageNet accuracy while holding robustness fixed (ablation) | Robust and standard pretrained models · CIFAR-10 | transfer accuracy: R^2 0.79; R^2 0.97 | Helped |
| Increase model width (ablation) | Robust pretrained models · Downstream classification tasks | transfer accuracy | Helped |
| Increase model width (ablation) | Standard pretrained models · Downstream classification tasks | transfer accuracy | No clear effect |
| Use task-selected robustness ε rather than one fixed ε for all tasks (ablation) | Robust pretrained models · CIFAR-10 versus CIFAR-100 transfer | accuracy: 3.0; 1.0 | Mixed |
| Downscale datasets to 32×32 before resizing to ImageNet scale (ablation) | Fixed-feature transfer · 12 classification datasets | transfer accuracy | No clear effect |
| Use adversarially robust rather than texture-invariant Stylized ImageNet pretraining | Transfer to standard datasets · Considered standard transfer tasks | accuracy | Helped |
| Use adversarially robust rather than standard ImageNet pretraining | Transfer evaluation or fine-tuning · Stylized ImageNet | accuracy | Helped |
Closest papers by meaning
- Adversarially robust transfer learning (ICLR 2020 · similarity 0.93)
- Adversarially-Trained Deep Nets Transfer Better: Illustration on Image Classification (ICLR 2021 · similarity 0.91)
- ImageNet Pre-training Also Transfers Non-robustness (AAAI 2023 · similarity 0.91)
- Does Robustness on ImageNet Transfer to Downstream Tasks? (CVPR 2022 · similarity 0.90)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.89)
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.