AI Paper Atlas

Research map / Adversarial robustness and security / Certified and empirical robustness

Non-Parametric Probabilistic Robustness: A Conservative Risk Estimator under Unknown Perturbation Distributions

Zheng Wang, Yi Zhang, Siddartha Khastgir, carsten maple, Xingyu Zhao

ICML 2026 · Adversarial robustness and security › Certified and empirical robustness · approach: Adversarial training recipes · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add learnable up-sampler to base GMM (K=3), label dependence (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%), lower is more conservative: 75.79 -> 69.08Improved
add learnable up-sampler, input-independent setting (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 74.99 -> 81.27Hurt
add learnable up-sampler, joint dependence (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 79.20 -> 78.59Improved
increase mixture components from K=3 to K=12, joint dependence (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 78.59 -> 73.11Improved
increase mixture components from K=7 to K=12, input-independent (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 73.25 -> 74.43Hurt
make perturbation distribution label dependent rather than independent (K=7, 16/255) (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 73.25 -> 64.88Improved
make perturbation input dependent rather than independent (K=7, 16/255) (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 73.25 -> 84.17Hurt
replace fixed Uniform perturbations with learned NPPR distributionPR evaluation · CIFAR-10 ResNet18 (16/255)robustness (%): 97.26 -> 76.27Improved
train estimator at larger radius (16/255 vs 4/255), evaluate at 32/255 (ablation)NPPR estimator · CIFAR-10 ResNet18NPPR (%): 59.79 -> 52.49Improved
use ResNet50 as GMM feature extractor instead of ResNet18, evaluating ResNet18 (ablation)NPPR estimator · CIFAR-10 (L_inf 16/255)NPPR (%): 76.27 -> 86.41Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy