Research map / Adversarial robustness and security / Attacks on text and graphs
Two Modalities Are Better Than One: Efficient Adversarial Purification via Multimodal Diffusion Models
Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Chao Li, Cesar F Caiafa, Junbin Gao, Qibin Zhao
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 |
|---|---|---|---|
| learn class-agnostic prompt instead of hand-crafted "a photo of a" template (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 70.29 -> 72.38 | Improved |
| learn class-agnostic prompt instead of hand-crafted template (ablation) | MultiDAP · CIFAR-100 AutoAttack l_inf | robust accuracy (%): 35.70 -> 38.15 | Improved |
| learn class-agnostic prompt instead of hand-crafted template (ablation) | MultiDAP · ImageNet-1K AutoAttack | robust accuracy (%): 38.65 -> 41.26 | Improved |
| add learned semantic prompt to unimodal likelihood maximization | LM · CIFAR-10 AutoAttack l_2 | robust accuracy (%): 75.00 -> 76.05 | Improved |
| increase purification steps from 5 to 50 (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 72.38 -> 53.75 | Hurt |
| increase purification steps from 1 to 5 (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 62.31 -> 72.38 | Improved |
| increase steps from 5 to 10 (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 72.38 -> 71.88 | No clear effect |
| use mid-range timestep range [400,600] instead of [100,200] (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 58.44 -> 72.38 | Improved |
| use very large timestep range [800,999] instead of [400,600] (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 72.38 -> 60.27 | Hurt |
| increase prompt length from 4 to 16 (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 71.19 -> 72.38 | Improved |
| increase prompt length from 16 to 32 (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 72.38 -> 72.04 | No clear effect |
| add proximity regularization lambda 0.9 vs none (ablation) | MultiDAP · CIFAR-10 AutoAttack l_inf | robust accuracy (%): 70.17 -> 72.38 | Improved |
Closest papers by meaning
- LoRID: Low-Rank Iterative Diffusion for Adversarial Purification (AAAI 2025 · similarity 0.89)
- Divide and Conquer: Heterogeneous Noise Integration for Diffusion-based Adversarial Purification (CVPR 2025 · similarity 0.89)
- Diffusion-based Adversarial Purification from the Perspective of the Frequency Domain (ICML 2025 · similarity 0.88)
- Diffusion Models Demand Contrastive Guidance for Adversarial Purification to Advance (ICML 2024 · similarity 0.88)
- Instant Adversarial Purification with Adversarial Consistency Distillation (CVPR 2025 · similarity 0.88)
- MimicDiffusion: Purifying Adversarial Perturbation via Mimicking Clean Diffusion Model (CVPR 2024 · 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.
