Research map / Adversarial robustness and security / Black-box and transfer attacks
Denoised Smoothing: A Provable Defense for Pretrained Classifiers
Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, J. Zico Kolter
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 a denoiser before the classifier | Pretrained ResNet-50 · ImageNet, ℓ2 radius 0.5 | certified top-1 accuracy: 4 -> 33 | Improved |
| Add a denoiser before the classifier | Pretrained ResNet-110 · CIFAR-10, ℓ2 radius 0.5 | certified accuracy: 3 -> 41 | Improved |
| Train the denoiser with Stab rather than MSE (ablation) | White-box ResNet-110 · CIFAR-10 | certified accuracy | Improved |
| Train the denoiser with Stab rather than Stab+MSE (ablation) | White-box ResNet-110 · CIFAR-10 | certified accuracy | Improved |
| Fine-tune an MSE denoiser with Stab (ablation) | White-box pretrained ResNet-18/34/50 · ImageNet | certified accuracy | Improved |
| Fine-tune an MSE denoiser with Stab using surrogate classifiers (ablation) | Black-box pretrained ResNet-18/34/50 · ImageNet | certified accuracy | Improved |
| Use 14 surrogate models rather than fewer surrogates to train the denoiser (ablation) | Black-box ResNet-110 · CIFAR-10, especially larger σ | certified accuracy | Improved |
| Use MSE rather than Stab (ablation) | Black-box CIFAR-10 denoiser at σ=0.50 · CIFAR-10, small ℓ2 radii | certified accuracy | Improved |
| Use Stab+MSE rather than MSE (ablation) | White-box ImageNet classifiers · ImageNet | certified accuracy | Improved |
| Use Stab+MSE rather than MSE (ablation) | Black-box pretrained classifiers · ImageNet | certified accuracy | Improved |
| Increase Monte Carlo samples from 100 to 1000 (ablation) | Azure Vision API with an MSE denoiser · ImageNet validation images | certified accuracy | Improved |
| Use the classification objective instead of the stability objective (ablation) | Denoisers on CIFAR-10, ImageNet, and vision APIs · CIFAR-10, ImageNet, and vision APIs | certified accuracy | No clear effect |
Closest papers by meaning
- Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers (NeurIPS 2019 · similarity 0.91)
- Certified Adversarial Robustness via Randomized Smoothing (ICML 2019 · similarity 0.89)
- (Certified!!) Adversarial Robustness for Free! (ICLR 2023 · similarity 0.89)
- Treatment of Statistical Estimation Problems in Randomized Smoothing for Adversarial Robustness (NeurIPS 2024 · similarity 0.87)
- Provably Cost-Sensitive Adversarial Defense via Randomized Smoothing (ICML 2025 · similarity 0.86)
- Mitigating the Curse of Dimensionality for Certified Robustness via Dual Randomized Smoothing (ICLR 2024 · similarity 0.86)
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.
