AI Paper Atlas

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

ICCV 2025 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training refinements · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add 100M synthetic samples to PGD-3 adversarial training (ablation)PGD-3 baseline · ImageNet-1K, AutoAttackadversarial accuracy: 55.82 -> 58.14Improved
use higher-quality REPA-generated synthetic data (50M) (ablation)PGD-3 with synthetic data · ImageNet-1K, AutoAttackadversarial accuracy: 58.14 -> 59.78Improved
switch architecture to ConvFormer (ablation)PGD-3 with synthetic data · ImageNet-1K, AutoAttackadversarial accuracy: 59.78 -> 63.14Improved
switch architecture from ConvFormer to CAFormer (ablation)PGD-3 with synthetic data · ImageNet-1K, AutoAttackadversarial accuracy: 63.14 -> 64.38Improved
add fine-tuning phase on real ImageNet data (ablation)CAFormer with synthetic data · ImageNet-1K, AutoAttackadversarial accuracy: 64.38 -> 65.32Improved
extend training from 300 to 700 epochs (ablation)CAFormer with synthetic data · ImageNet-1K, AutoAttackadversarial accuracy: 65.32 -> 66.58Improved
use Stable-PGD (adaptive sign scaling) instead of PGD on a small stable model (ablation)ConvNeXt-B 300 epochs · ImageNet-1K, AutoAttackrobust 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 curvestraining stabilityNo clear effect
use StableAdamW, lower beta2 or stronger augmentation to suppress loss oscillation (ablation)Large-model adversarial training · ImageNet-1K training curvestraining stabilityNo clear effect
randomize PGD step size by factor from U(0.5, 1) (ablation)PGD-3 · ImageNet-1K training curvesloss oscillationMixed
use purely convolutional ConvFormer instead of attention-based CAFormer (ablation)Large-model adversarial training · ImageNet-1K training curvesloss oscillationNo clear effect
scale to CAFormer-XXL48 with 50M synthetic samples versus AdvXL ViT-g with 1B external samplesCAFormer · ImageNet-1K, AutoAttack linfrobust accuracy: 71.0 -> 71.54Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy