Non-parametric probabilistic robustness NPPR: a conservative risk estimator under unknown perturbation distributions

The study proposes the non-parametric probabilistic robustness (NPPR) metric, which learns the perturbation distribution from data under an unknown perturbation distribution and gives a conservative risk estimate. It builds an estimator based on a Gaussian mixture model (GMM), covering both input-dependent and input-independent perturbation scenarios; on CIFAR-10, CIFAR-100, Tiny ImageNet and ResNet18/50, WideResNet50, VGG16, the PR estimate of NPPR is lower than results when a common perturbation distribution is assumed.