Verification of machine learning based cyber-physical systems: a comparative study

Abstract

In this paper, we conduct a comparison of the existing formal methods for verifying the safety of cyber-physical systems with machine learning based controllers. We focus on a particular form of machine learning based controller, namely a classifier based on multiple neural networks, the architecture of which is particularly interesting for embedded applications. We compare both exact and approximate verification techniques, based on several real-world benchmarks such as a collision avoidance system for unmanned aerial vehicles.