Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Robustification via Text-to-Image Diffusion Models
Daewon Choi, Jongheon Jeong, Huiwon Jang, Jinwoo Shin
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 self-adaptation (T2I personalization and classifier fine-tuning) (ablation) | CLIP + denoised smoothing · ImageNet, eps=0.5 | certified accuracy (%): 29.6 -> 34.2 | Improved |
| add self-adaptation (T2I personalization and classifier fine-tuning) (ablation) | ResNet-50 + denoised smoothing · ImageNet ResNet-50, eps=1.0 | robust accuracy (%): 44.2 -> 47.8 | Improved |
| adapt text-to-image diffusion model only (without classifier adaptation) (ablation) | CLIP + denoised smoothing · ImageNet, eps=0.5 | certified accuracy (%): 29.6 -> 31.8 | Improved |
| also adapt classifier on top of T2I adaptation (ablation) | CLIP + denoised smoothing · ImageNet, eps=0.5 | certified accuracy (%): 31.8 -> 34.2 | Improved |
| increase timestep correction factor k above 1 for super-resolution denoiser (ablation) | denoiser · ImageNet subset denoised accuracy | CLIP accuracy of denoised images | Improved |
| add classifier-guided regularization (lambda=0.01) in personalization (ablation) | T2I personalization · ImageNet, sigma=0.5 | ACR: 0.358 -> 0.394 | Improved |
| increase reference set from 1-shot to 4-shot per class (ablation) | self-adaptation · ImageNet, sigma=0.5 | ACR: 0.409 -> 0.428 | Improved |
| increase number of noise samples n from 25 to 400 (ablation) | denoised smoothing inference · ImageNet, eps=1.0 | robust vs clean accuracy: +9.2% robust, -3.8% clean, 8.0x inference time | Mixed |
| use data-free denoised smoothing instead of Mao et al. adversarial fine-tuning on ImageNet | CLIP · EuroSAT, eps=1.0 | robust accuracy (%): 0.0 -> 28.2 | Improved |
Closest papers by meaning
- Robust Classification via a Single Diffusion Model (ICML 2024 · similarity 0.91)
- Robustness in Both Domains: CLIP Needs a Robust Text Encoder (NeurIPS 2025 · similarity 0.90)
- Adaptive Median Smoothing: Adversarial Defense for Unlearned Text-to-Image Diffusion Models at Inference Time (ICML 2025 · similarity 0.89)
- CLIP is Strong Enough to Fight Back: Test-time Counterattacks towards Zero-shot Adversarial Robustness of CLIP (CVPR 2025 · similarity 0.89)
- Efficient Image-to-Image Diffusion Classifier for Adversarial Robustness (AAAI 2025 · similarity 0.89)
- CAT: Contrastive Adversarial Training for Evaluating the Robustness of Protective Perturbations in Latent Diffusion Models (ICML 2025 · similarity 0.89)
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.
