Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarially robust transfer learning
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 |
|---|---|---|---|
| Transfer a robust feature extractor and train only the target classification layer | Robust CIFAR-100 model · CIFAR-10 transferred from CIFAR-100 | PGD-20 accuracy: 17.70% | Helped |
| Use a naturally trained rather than robust source feature extractor | Frozen-feature transfer to CIFAR-10 · CIFAR-10 transferred from CIFAR-100 | validation accuracy: 72.05% -> 83.05% | Helped |
| Transfer robust features between more similar source and target datasets | Frozen-feature transfer · CIFAR-100 split by classes | PGD-20 accuracy: 15.86% vs 11.63% | Helped |
| Use robust ImageNet features instead of training a robust u-ResNet-50 on the target | CIFAR-10+ · CIFAR-10+ | validation accuracy: 82.00% -> 88.33% | Helped |
| Use a frozen robust ImageNet feature extractor rather than a natural ImageNet feature extractor | ImageNet-to-CIFAR transfer · CIFAR-10+ | PGD-20 accuracy: 0.01% -> 22.66% | Helped |
| Add hidden layers to the classifier on top of a robust feature extractor (ablation) | ImageNet-to-CIFAR-100 MLP transfer · CIFAR-100 | PGD-20 accuracy: robustness ... improves | Helped |
| Add hidden layers to the classifier on top of a robust feature extractor (ablation) | ImageNet-to-CIFAR-100 MLP transfer · CIFAR-100 | validation accuracy | Hurt |
| Add dropout and batch normalization to the deeper classifier (ablation) | ImageNet-to-CIFAR-100 MLP transfer · CIFAR-100 | validation accuracy | Helped |
| Fine-tune end-to-end with penultimate-feature distillation rather than freeze the feature extractor | Robust CIFAR-100 model transferred to CIFAR-10 · CIFAR-10 | validation accuracy: 72.05% -> 79.35% | Helped |
| Increase the feature-distillation weight from 1e-7 to 0.005 (ablation) | LwF fine-tuning from robust CIFAR-100 to CIFAR-10 · CIFAR-10 | PGD-20 accuracy: 0.61% -> 17.61% | Helped |
| Increase the feature-distillation weight from 1e-7 to 0.005 (ablation) | LwF fine-tuning from robust CIFAR-100 to CIFAR-10 · CIFAR-10 | validation accuracy: 89.07% -> 79.35% | Hurt |
| Apply LwF feature distillation when transferring from robust CIFAR-100 to CIFAR-100 | Robust CIFAR-100 model · CIFAR-100 | validation accuracy: 59.87% -> 61.71% | Helped |
Closest papers by meaning
- Do Adversarially Robust ImageNet Models Transfer Better? (NeurIPS 2020 · similarity 0.93)
- Adversarially Robust Hypothesis Transfer Learning (ICML 2024 · similarity 0.89)
- Adversarially Robust Multi-task Representation Learning (NeurIPS 2024 · similarity 0.89)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.88)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · 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.