Research map / Adversarial robustness and security / Certified and empirical robustness
Scaling and Taming Adversarial Training with Synthetic Data
Juntao Wu, Xianting Huang, Yu Chen, Shuai Pang, Ke Wang
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 |
|---|---|---|---|
| add 100M synthetic samples to PGD-3 adversarial training (ablation) | PGD-3 baseline · ImageNet-1K, AutoAttack | adversarial accuracy: 55.82 -> 58.14 | Improved |
| use higher-quality REPA-generated synthetic data (50M) (ablation) | PGD-3 with synthetic data · ImageNet-1K, AutoAttack | adversarial accuracy: 58.14 -> 59.78 | Improved |
| switch architecture to ConvFormer (ablation) | PGD-3 with synthetic data · ImageNet-1K, AutoAttack | adversarial accuracy: 59.78 -> 63.14 | Improved |
| switch architecture from ConvFormer to CAFormer (ablation) | PGD-3 with synthetic data · ImageNet-1K, AutoAttack | adversarial accuracy: 63.14 -> 64.38 | Improved |
| add fine-tuning phase on real ImageNet data (ablation) | CAFormer with synthetic data · ImageNet-1K, AutoAttack | adversarial accuracy: 64.38 -> 65.32 | Improved |
| extend training from 300 to 700 epochs (ablation) | CAFormer with synthetic data · ImageNet-1K, AutoAttack | adversarial accuracy: 65.32 -> 66.58 | Improved |
| use Stable-PGD (adaptive sign scaling) instead of PGD on a small stable model (ablation) | ConvNeXt-B 300 epochs · ImageNet-1K, AutoAttack | robust accuracy: within +/-0.2% | No clear effect |
| replace PGD with APGD in training to remove loss oscillation (ablation) | Large-model adversarial training · ImageNet-1K training curves | training stability | No clear effect |
| use StableAdamW, lower beta2 or stronger augmentation to suppress loss oscillation (ablation) | Large-model adversarial training · ImageNet-1K training curves | training stability | No clear effect |
| randomize PGD step size by factor from U(0.5, 1) (ablation) | PGD-3 · ImageNet-1K training curves | loss oscillation | Mixed |
| use purely convolutional ConvFormer instead of attention-based CAFormer (ablation) | Large-model adversarial training · ImageNet-1K training curves | loss oscillation | No clear effect |
| scale to CAFormer-XXL48 with 50M synthetic samples versus AdvXL ViT-g with 1B external samples | CAFormer · ImageNet-1K, AutoAttack linf | robust accuracy: 71.0 -> 71.54 | Improved |
Closest papers by meaning
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.90)
- Revisiting Adversarial Training at Scale (CVPR 2024 · similarity 0.89)
- Scaling provable adversarial defenses (NeurIPS 2018 · similarity 0.89)
- Scaling Laws of Synthetic Images for Model Training ... for Now (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)
- ROME is Forged in Adversity: Robust Distilled Datasets via Information Bottleneck (ICML 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.
