Research map / Label-efficient and robust learning / Recognition and long-tail
Adversarially-Trained Deep Nets Transfer Better: Illustration on Image Classification
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 |
|---|---|---|---|
| Adversarially pre-train the ImageNet source model instead of training it naturally | ResNet-50 transfer learning · Six target datasets, especially with limited target data | test accuracy: 48.49 -> 51.89 (CIFAR-10, 100 images); 23.5 -> 37.9 (SVHN, 100 images) | Helped |
| Adversarially pre-train the source model with ℓ2 ≤ 3 rather than ℓ∞ constraints (ablation) | ImageNet-to-target transfer · Target datasets, fine-tuning three convolutional blocks | test accuracy | Helped |
| Use a smaller adversarial perturbation constraint (ablation) | ImageNet-to-CIFAR-10 and ImageNet-to-CIFAR-100 transfer · CIFAR-10 and CIFAR-100 | test accuracy | Helped |
| Fine-tune one or three convolutional blocks rather than zero or nine (ablation) | Transferred ResNet-50 · Six target datasets | test accuracy | Helped |
| Use fewer fine-tuned blocks (ablation) | ImageNet-to-CIFAR transfer · CIFAR-10 and CIFAR-100 | test accuracy | Helped |
| Use PGD(20) rather than PGD(1) for adversarial source training (ablation) | ImageNet source model · Six target datasets | test accuracy | Helped |
| Use targeted PGD perturbations rather than Gaussian noise during source training (ablation) | ImageNet source model · Six target datasets | test accuracy | Helped |
| Add Gaussian noise during source training rather than use natural training (ablation) | ImageNet source model · Six target datasets | test accuracy | Helped |
| Adversarially train the source model rather than train it naturally | ImageNet-to-Stylized ImageNet transfer · Stylized ImageNet test | test accuracy: 11.4 -> 20.1 | Helped |
| Adversarially train the source model rather than train it naturally | ImageNet-to-Stylized ImageNet transfer · Stylized ImageNet after fine-tuning | test accuracy: 36.1 -> 64.2 | Helped |
| Adversarially train the source model rather than train it naturally | Caltech101 transfer · Low-Res | test accuracy: 79.1 -> 83.7 | Helped |
| Adversarially train the source model rather than train it naturally | Caltech101 transfer · Low-Pass | test accuracy: 80.7 -> 85.3 | Helped |
Closest papers by meaning
- Adversarial Training Helps Transfer Learning via Better Representations (NeurIPS 2021 · similarity 0.92)
- Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · similarity 0.91)
- Transferable Adversarial Training: A General Approach to Adapting Deep Classifiers (ICML 2019 · similarity 0.89)
- What is being transferred in transfer learning? (NeurIPS 2020 · similarity 0.89)
- Learning to Transform Dynamically for Better Adversarial Transferability (CVPR 2024 · similarity 0.87)
- Boosting Adversarial Transferability through Augmentation in Hypothesis Space (CVPR 2025 · 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.