A Zonotopic Dempster-Shafer Approach to the Quantitative Verification of Neural Networks

Abstract The reliability and usefulness of verification depend on the ability to represent appropriately the uncertainty. Most existing work on neural network verification relies on the hypothesis of either set-based or probabilistic information on the inputs. In this work, we rely on the framework of imprecise probabilities, specifically p-boxes, to propose a quantitative verification of ReLU neural networks, which can account for both probabilistic information and epistemic uncertainty on inputs. On classical benchmarks, including the ACAS Xu examples, we demonstrate that our approach improves the tradeoff between tightness and efficiency compared to related work on probabilistic network verification, while handling much more general classes of uncertainties on the inputs and providing fully guaranteed results.