This paper presents the Neural Network Verification (NNV) software tool, a set-based verification framework for deep neural networks (DNNs) and learning-enabled cyber-physical systems (CPS). The crux of NNV is a collection of reachability algorithms that make use of a variety of set representations, such as polyhedra, star sets, zonotopes, and abstract-domain representations. NNV supports both exact (sound and complete) and over-approximate (sound) reachability algorithms for verifying safety and robustness properties of feed-forward neural networks (FFNNs) with various activation functions. For learning-enabled CPS, such as closed-loop control systems incorporating neural networks, NNV provides exact and over-approximate reachability analysis schemes for linear plant models and FFNN controllers with piecewise-linear activation functions, such as ReLUs. For similar neural network control systems (NNCS) that instead have nonlinear plant models, NNV supports overapproximate analysis by combining the star set analysis used for FFNN controllers with zonotope-based analysis for nonlinear plant dynamics building on CORA. We evaluate NNV using two real-world case studies: the first is safety verification of ACAS Xu networks, and the second deals with the safety verification of a deep learning-based adaptive cruise control system.

No abstract

No abstract

No abstract

This benchmark suite is composed of nine examples of large-scale linear systems, ranging in dimensionality in the tens to the low thousands. The benchmarks are derived from diverse fields such as civil engineering and robotics, and are based on similar existing test sets for model-order reduction algorithms in control and numerical analysis. Each example is provided in the SpaceEx XML model format as single-mode hybrid automaton and are compatible with the HyST model transformation tool to support analysis in other verification tools. Some preliminary reachability analysis results for some of the smaller examples (on the order of tens of dimensions) are presented using SpaceEx.