This paper presents a quantitative robustness measure, instead of the binary notion of robustness for evaluating Neural Networks, we also consider input feature interdependence in defining neighborhoods of an input point. Each point’s influence in the neighborhood towards the score is determined by a probability distribution, here Gaussian, in contrast to the prior uniform distribution approach. We offer two methods for computation of our robustness score: 1) an approximate method and 2) an abstract method that computes an interval containing the true score. Experiments show that the concrete method is comparable to and notably faster than the abstract method. We apply these methods to assess the robustness of neural networks trained on MNIST and Higgs datasets.