Research map / Adversarial robustness and security / Certified and empirical robustness
(Certified!!) Adversarial Robustness for Free!
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 |
|---|---|---|---|
| combine a pretrained diffusion denoiser with an off-the-shelf classifier for randomized smoothing | ImageNet classification · ImageNet | certified top-1 accuracy at ε=0.5 (%): 57.0 -> 71.1 | Helped |
| use diffusion denoised smoothing rather than the strongest listed baseline | ImageNet classification · ImageNet, ε=3.0 | certified top-1 accuracy (%): 23.4 -> 13.1 | Hurt |
| use diffusion denoised smoothing with an off-the-shelf ViT | CIFAR-10 classification · CIFAR-10, ε=0.25 | certified accuracy (%): 74.9 -> 76.7 | Helped |
| replace ViT with CIFAR-10-only Wide-ResNet (ablation) | Diffusion Denoised Smoothing · CIFAR-10, ε=0.5 | certified accuracy (%): 63.0 -> 55.7 | Hurt |
| finetune Wide-ResNet on noised-then-denoised training images (ablation) | Diffusion Denoised Smoothing · CIFAR-10, ε=0.5 | certified accuracy (%): 55.7 -> 63.8 | Helped |
| finetune ViT on noised-then-denoised training images (ablation) | Diffusion Denoised Smoothing · CIFAR-10, ε=0.5 | certified accuracy (%): 63.0 -> 65.5 | Helped |
| train the diffusion denoiser on three noise levels instead of only σ=0.5 (ablation) | one-shot denoising with ViT · CIFAR-10 denoised images, evaluation σ=0.5 | clean accuracy: 60.1 -> 68.1 | Helped |
| train the denoiser on a broad standard noise range instead of only σ=0.5 (ablation) | one-shot denoising with ViT · CIFAR-10 denoised images, evaluation σ=0.5 | clean accuracy: 60.1 -> 72.3 | Helped |
| train the denoiser on the standard broad range rather than [0, 1.0] (ablation) | one-shot denoising with ViT · CIFAR-10 denoised images, evaluation σ=1.0 | clean accuracy: 46.0 -> 44.8 | Hurt |
| use deterministic multi-step EDM sampling instead of one-shot denoising (ablation) | diffusion denoising with ViT · CIFAR-10 denoised images, σ=1.0 | clean accuracy: 50.8 -> 54.0 | Helped |
Closest papers by meaning
- Certified Adversarial Robustness via Randomized Smoothing (ICML 2019 · similarity 0.92)
- Improving l1-Certified Robustness via Randomized Smoothing by Leveraging Box Constraints (ICML 2023 · similarity 0.90)
- Robust Representation Consistency Model via Contrastive Denoising (ICLR 2025 · similarity 0.89)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.89)
- Robust Classification via a Single Diffusion Model (ICML 2024 · 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.