Non-Parametric Probabilistic Robustness: A Conservative Risk Estimator under Unknown Perturbation Distributions
Researchers have introduced Non-Parametric Probabilistic Robustness (NPPR), a new metric for evaluating the robustness of deep learning models. Unlike previous methods that assume a known perturbation distribution, NPPR learns this distribution directly from data, offering a more practical assessment under uncertainty. An NPPR estimator using Gaussian Mixture Models was developed, and theoretical analyses show its relationship to existing adversarial and probabilistic robustness metrics. Experiments on standard datasets and various model architectures demonstrate that NPPR provides more conservative robustness estimates. AI
IMPACT Introduces a more practical metric for assessing model safety and reliability under unknown data perturbations.