Research map / Adversarial robustness and security / Certified and empirical robustness
Fast is better than free: Revisiting adversarial training
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 |
|---|---|---|---|
| Add random initialization to FGSM adversarial training instead of zero initialization (ablation) | FGSM adversarial training · CIFAR10, ε=8/255 | PGD accuracy: 0.00% -> 44.01% | Helped |
| Initialize from the previous minibatch perturbation instead of zero initialization (ablation) | FGSM adversarial training · CIFAR10, ε=8/255 | PGD accuracy: 0.00% -> 42.37% | Helped |
| Increase step size to α=10/255 (ablation) | FGSM with random initialization · CIFAR10, ε=8/255 | PGD accuracy: 44.01% -> 46.06% | Helped |
| Increase step size to α=16/255 (ablation) | FGSM with random initialization · CIFAR10, ε=8/255 | PGD accuracy: 44.01% -> 0.00% | Hurt |
| Add early stopping (ablation) | FGSM with zero initialization · CIFAR10, ε=8/255 | PGD accuracy: 0.00% -> 38.86% | Helped |
| Add early stopping (ablation) | FGSM with α=16/255 · CIFAR10, ε=8/255 | PGD accuracy: 0.00% -> 40.38% | Helped |
| Replace the original R+FGSM initialization with uniform initialization (ablation) | R+FGSM adversarial training · MNIST | robust accuracy: 34.58± 36.06% -> 72.92 ± 10.40% | Helped |
| Use a full step size instead of the original smaller step size (ablation) | R+FGSM adversarial training · MNIST | robust accuracy: 34.58± 36.06% -> 26.53 ± 32.48% | Hurt |
| Combine uniform initialization with a full step size (ablation) | R+FGSM adversarial training · MNIST | robust accuracy: 34.58± 36.06% -> 86.21 ± 00.75% | Helped |
| Use mixed precision (ablation) | FGSM adversarial training with DAWNBench techniques · ImageNet | total training time (hours): 14.59 -> 12.14 | Helped |
| Use mixed precision (ablation) | Free adversarial training · ImageNet | total training time (hours): 52.20 -> 38.76 | Helped |
| Apply DAWNBench optimizations to Free training for 15 epochs (ablation) | Free adversarial training (m=3) · ImageNet, ε=4/255 | PGD+10 accuracy: 31.08% -> 22.18% | Hurt |
Closest papers by meaning
- Adversarial training for free! (NeurIPS 2019 · similarity 0.92)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.91)
- Bag of Tricks for Adversarial Training (ICLR 2021 · similarity 0.91)
- Adversarial Initialization with Universal Adversarial Perturbation: A New Approach to Fast Adversarial Training (AAAI 2024 · similarity 0.90)
- Finding Actual Descent Directions for Adversarial Training (ICLR 2023 · similarity 0.90)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · 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.