Research map / Adversarial robustness and security / Certified and empirical robustness
Better Diffusion Models Further Improve 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 |
|---|---|---|---|
| Replace DDPM-generated training data with EDM-generated data | WRN-70-16 adversarial training · CIFAR-10, ℓ∞ ε=8/255 | AutoAttack accuracy (%): 66.11 -> 70.69 | Helped |
| Use EDM-generated data instead of the strongest listed baseline data | WRN-70-16 adversarial training · CIFAR-10, ℓ2 ε=128/255 | AutoAttack accuracy (%): 82.32 -> 84.86 | Helped |
| Replace the strongest listed baseline with EDM-generated data | WRN-70-16 adversarial training · CIFAR-100, ℓ∞ ε=8/255 | AutoAttack accuracy (%): 36.88 -> 42.67 | Helped |
| Use 50M EDM-generated images | WRN-28-10 adversarial training · SVHN, ℓ∞ ε=8/255 | AutoAttack accuracy (%): 61.09 -> 64.01 | Helped |
| Use ImageNet-pretrained EDM-generated data instead of class-conditional DDPM data | WRN-28-10 adversarial training · TinyImageNet, ℓ∞ ε=8/255 | AutoAttack accuracy (%): 26.66 -> 31.30 | Helped |
| Replace common augmentation with Cutout (ablation) | EDM adversarial training · CIFAR-10 | AutoAttack accuracy (%): 63.35 -> 63.30 | No clear effect |
| Replace common augmentation with CutMix (ablation) | EDM adversarial training · CIFAR-10 | AutoAttack accuracy (%): 63.35 -> 62.81 | Hurt |
| Replace common augmentation with AutoAugment (ablation) | EDM adversarial training · CIFAR-10 | AutoAttack accuracy (%): 63.35 -> 62.86 | Hurt |
| Replace common augmentation with RandAugment (ablation) | EDM adversarial training · CIFAR-10 | AutoAttack accuracy (%): 63.35 -> 63.12 | Hurt |
| Replace common augmentation with IDBH (ablation) | EDM adversarial training · CIFAR-10 | AutoAttack accuracy (%): 63.35 -> 63.24 | Hurt |
| Use class-conditional rather than unconditional generation (ablation) | EDM data generation for adversarial training · CIFAR-10 | AutoAttack accuracy (%): 62.45 -> 63.35 | Helped |
| Increase generated data from no synthetic images to 1M images (ablation) | WRN-28-10 adversarial training · CIFAR-10 | AutoAttack accuracy (%): 45.29 -> 63.50 | Helped |
Closest papers by meaning
- Improved Diffusion-based Generative Model with Better Adversarial Robustness (ICLR 2025 · similarity 0.90)
- CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models (ICML 2025 · similarity 0.89)
- Robust Classification via a Single Diffusion Model (ICML 2024 · similarity 0.89)
- ACT-Diffusion: Efficient Adversarial Consistency Training for One-step Diffusion Models (CVPR 2024 · similarity 0.89)
- Structure-Guided Adversarial Training of Diffusion Models (CVPR 2024 · similarity 0.88)
- Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective (ICLR 2025 · similarity 0.88)
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.