Research map / Adversarial robustness and security / Certified and empirical robustness
Phase Transition from Clean Training to Adversarial Training
Yue Xing, Qifan Song, Guang Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Increase attack strength eps from 0 to 0.5/255 | Adversarial training PreActResNet18 · CIFAR-10 | adversarial testing loss: 0.1837 -> 0.2937 | Hurt |
| Increase attack strength eps from 0.5/255 to 2/255 | Adversarial training · CIFAR-10 | adversarial testing accuracy: 0.8977 -> 0.8101 | Hurt |
| Increase eps beyond eps* (0.5, then 2, 8) | Adversarial training · CIFAR-10 connectivity | low-loss region overlap with clean model | Hurt |
| Use FGM instead of PGD at eps below eps* | FGSM adversarial training · CIFAR-10 | catastrophic overfitting: 0.5/255 vs 8/255 | No clear effect |
| Reduce training size from 50K to 10K | Adversarial training PreActResNet18 · CIFAR-10 | eps*: 0.5/255 -> 3.3/255 | Improved |
| Reduce training size from 50K to 25K | Adversarial training PreActResNet18 · CIFAR-100 | eps*: 1.0/255 -> 4.0/255 | Improved |
| Widen MNIST hidden layer from 16 to 128 nodes | MNIST CNN adversarial training · MNIST | eps*: 26.1/255 -> 48.0/255 | Improved |
| Widen WideResNet34 from width 1 to 10 | Adversarial training · CIFAR-10 | eps*: 1.1/255 -> 1.1/255 | No clear effect |
| Replace vanilla adversarial loss with TRADES | PreActResNet18 · CIFAR-10 | eps*: 0.5/255 -> 0.7/255 | Improved |
| Replace vanilla adversarial loss with MART | PreActResNet18 · CIFAR-10 | eps*: 0.5/255 -> 0.8/255 | Improved |
| Estimate eps* by adversarial fine-tuning with PGD, 1 to 3 runs | Extrapolation algorithm · CIFAR-10 n=10000 Linf | eps* estimate (true 3.3/255): 4.7/255 -> 2.5/255 | Mixed |
| Estimate eps* with FGM 3 runs instead of PGD 3 runs | Extrapolation algorithm · CIFAR-10 n=50000 Linf | eps* estimate (true 0.5/255): 0.52/255 -> 0.59/255 | Hurt |
Closest papers by meaning
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · similarity 0.89)
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.88)
- Efficient Robust Training via Backward Smoothing (AAAI 2022 · similarity 0.88)
- Over-parameterized Adversarial Training: An Analysis Overcoming the Curse of Dimensionality (NeurIPS 2020 · similarity 0.88)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.87)
- On the Scalability of Certified Adversarial Robustness with Generated Data (NeurIPS 2024 · similarity 0.86)
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.
