Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial Training Reduces Information and Improves Transferability
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 rather than standard pretraining on CIFAR-10 | ResNet-50 linear transfer · MNIST | accuracy (%): 64.17 -> 88.11 | Helped |
| Use adversarial rather than standard pretraining on CIFAR-10 | ResNet-50 linear transfer · CIFAR-100 | accuracy (%): 35.76 -> 44.92 | Helped |
| Use adversarial rather than standard pretraining on CIFAR-100 | ResNet-50 linear transfer · CIFAR-10 | accuracy (%): 80.18 -> 74.47 | Hurt |
| Use adversarial rather than standard pretraining on ImageNet | ResNet-50 linear transfer · SVHN | accuracy (%): 50.46 -> 76.90 | Helped |
| Use adversarial rather than standard pretraining on ImageNet | ResNet-50 linear transfer · Birds | accuracy (%): 53.58 -> 35.91 | Hurt |
| Use adversarial rather than standard pretraining on CIFAR-10 | ResNet-50 linear transfer · CINIC-IMAGENET | accuracy (%): 84 -> 80 | Hurt |
| Fine-tune the linear classifier and batch norm across the entire network instead of only the classifier and its preceding batch norm (ablation) | ResNet-50 pretrained on CIFAR-10 · CIFAR-100 | accuracy (%): 44.92 -> 61.92 | Helped |
| Fine-tune the entire network instead of only the classifier and its preceding batch norm (ablation) | ResNet-50 pretrained on CIFAR-10 · CIFAR-100 | accuracy (%): 44.92 -> 78.85 | Helped |
| Fine-tune the linear classifier and batch norm across the entire network instead of only the classifier and its preceding batch norm (ablation) | ResNet-50 pretrained on ImageNet · Aircraft | accuracy (%): 33.81 -> 37.89 | Helped |
| Fine-tune the entire network instead of only the classifier and its preceding batch norm (ablation) | ResNet-50 pretrained on ImageNet · Aircraft | accuracy (%): 33.81 -> 67.15 | Helped |
| Use adversarial rather than standard pretraining on ImageNet | ResNet-50 transfer with full-network fine-tuning · Aircraft | accuracy (%): 71.32 -> 67.15 | Hurt |
| Inject effective noise once at the start of inversion | Image reconstruction · Image inversion | inversion loss | Helped |
Closest papers by meaning
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Adversarial Training Helps Transfer Learning via Better Representations (NeurIPS 2021 · similarity 0.89)
- Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · similarity 0.89)
- Adversarially-Trained Deep Nets Transfer Better: Illustration on Image Classification (ICLR 2021 · similarity 0.89)
- Robust Models are less Over-Confident (NeurIPS 2022 · similarity 0.89)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.87)
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.