Research map / Adversarial robustness and security / Attacks on text and graphs
Diffusion Models for Adversarial Purification
Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, Animashree Anandkumar
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 |
|---|---|---|---|
| Use DiffPure instead of the strongest baseline without extra data | WideResNet-28-10 · CIFAR-10, AutoAttack ℓ∞ ε=8/255 | robust accuracy: 65.24 -> 70.64±0.39 | Improved |
| Use DiffPure instead of the strongest baseline without extra data | WideResNet-70-16 · CIFAR-10, AutoAttack ℓ∞ ε=8/255 | robust accuracy: 66.60 -> 71.29±0.55 | Improved |
| Use DiffPure instead of the strongest baseline without extra data | WideResNet-28-10 · CIFAR-10, AutoAttack ℓ2 ε=0.5 | robust accuracy: 78.32 -> 78.58±0.40 | Improved |
| Use DiffPure instead of the strongest baseline without extra data | WideResNet-70-16 · CIFAR-10, AutoAttack ℓ2 ε=0.5 | robust accuracy: 80.86 -> 80.60±0.57 | Mixed |
| Use DiffPure instead of the strongest baseline | ResNet-50 · ImageNet, AutoAttack ℓ∞ ε=4/255 | robust accuracy: 37.89 -> 40.93±1.96 | Improved |
| Use DiffPure instead of the baseline | WideResNet-50-2 · ImageNet, AutoAttack ℓ∞ ε=4/255 | robust accuracy: 39.25 -> 44.39±0.95 | Improved |
| Use DiffPure instead of the baseline | DeiT-S · ImageNet, AutoAttack ℓ∞ ε=4/255 | robust accuracy: 35.50 -> 43.18±1.27 | Improved |
| Use DiffPure instead of the strongest unseen-threat baseline | ResNet-50 · CIFAR-10, AutoAttack ℓ∞ ε=8/255 | robust accuracy: 40.0 -> 70.0±1.2 | Improved |
| Use DiffPure instead of the strongest unseen-threat baseline | ResNet-50 · CIFAR-10, AutoAttack ℓ2 ε=1 | robust accuracy: 34.9 -> 70.9±0.6 | Improved |
| Use DiffPure instead of the strongest purification baseline | Eyeglasses attribute classifier · CelebA-HQ, BPDA+EOT ℓ∞ ε=16/255 | robust accuracy: 75.00 -> 90.63±1.10 | Improved |
| Use DiffPure instead of DSM+LD | WideResNet-28-10 · CIFAR-10, BPDA+EOT ℓ∞ ε=8/255 | robust accuracy: 70.01 -> 81.40±0.16 | Improved |
| Replace VP-SDE sampling with VP-ODE sampling (ablation) | DiffPure · CIFAR-10, AutoAttack ℓ∞ ε=8/255 | robust accuracy: 70.64±0.39 -> 39.86±0.98 | Hurt |
Closest papers by meaning
- LoRID: Low-Rank Iterative Diffusion for Adversarial Purification (AAAI 2025 · similarity 0.93)
- Diffusion Models Demand Contrastive Guidance for Adversarial Purification to Advance (ICML 2024 · similarity 0.93)
- Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification (CVPR 2025 · similarity 0.92)
- MimicDiffusion: Purifying Adversarial Perturbation via Mimicking Clean Diffusion Model (CVPR 2024 · similarity 0.92)
- ADBM: Adversarial Diffusion Bridge Model for Reliable Adversarial Purification (ICLR 2025 · similarity 0.92)
- Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain (ICML 2025 · similarity 0.92)
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.
