Research map / Adversarial robustness and security / Certified and empirical robustness
Attacks Which Do Not Kill Training Make Adversarial Learning Stronger
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 early-stopped PGD friendly adversarial training instead of DAT | WRN-32-10 · CIFAR-10 | Natural accuracy: 85.03 -> 89.34 ± 0.221 | Helped |
| Use early-stopped PGD friendly adversarial training instead of DAT | WRN-32-10 · CIFAR-10 | FGSM accuracy: 63.53 -> 65.52 ± 0.355 | Helped |
| Use early-stopped PGD friendly adversarial training instead of DAT | WRN-32-10 · CIFAR-10 | PGD-20 accuracy: 48.70 -> 46.13 ± 0.409 | Hurt |
| Increase FAT training perturbation bound from 8/255 to 16/255 | WRN-32-10 FAT · CIFAR-10 | PGD-20 accuracy: 46.13 ± 0.409 -> 49.86 ± 0.328 | Helped |
| Increase FAT training perturbation bound from 8/255 to 16/255 | WRN-32-10 FAT · CIFAR-10 | Natural accuracy: 89.34 ± 0.221 -> 87.00 ± 0.203 | Hurt |
| Use early-stopped PGD in TRADES | WRN-34-10, TRADES β=1.0 · CIFAR-10 | Natural accuracy: 88.64 -> 89.94 ± 0.303 | Helped |
| Use early-stopped PGD in TRADES | WRN-34-10, TRADES β=1.0 · CIFAR-10 | FGSM accuracy: 56.38 -> 61.00 ± 0.418 | Helped |
| Use early-stopped PGD in TRADES | WRN-34-10, TRADES β=6.0 · CIFAR-10 | PGD-20 accuracy: 56.61 -> 55.98 ± 0.209 | Hurt |
| Increase FAT for TRADES training perturbation bound from 8/255 to 16/255 | WRN-34-10, TRADES β=6.0 · CIFAR-10 | PGD-20 accuracy: 55.98 ± 0.209 -> 57.12 ± 0.233 | Helped |
| Increase the early-stop step τ (ablation) | FAT · CIFAR-10 | Natural test accuracy | Hurt |
| Increase the early-stop step τ (ablation) | FAT · CIFAR-10 | Robust test accuracy | Mixed |
| Use smaller τ in early-stopped PGD (ablation) | FAT · Adversarial training | Backward propagations per example | Helped |
Closest papers by meaning
- Improving Adversarial Robustness by Putting More Regularizations on Less Robust Samples (ICML 2023 · similarity 0.91)
- More Data Can Expand The Generalization Gap Between Adversarially Robust and Standard Models (ICML 2020 · similarity 0.90)
- Formulating Robustness Against Unforeseen Attacks (NeurIPS 2022 · similarity 0.90)
- Efficient Robust Training via Backward Smoothing (AAAI 2022 · similarity 0.90)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation (NeurIPS 2024 · 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.