We apply differential dynamic (refinement) logic to a case study on neural highway control. After modelling the abstract system we use VerSAILLE and the NCubeV tool [NeurIPS24] to (dis)prove the safety of concrete neural networks. Along the way we uncover numerous interesting results on the highway-env simulation environment including inconsistencies between the provided specification and the actual simulation.
10. Jun 2025
We introduce a new abstract domain for differential verification using Zonotopes and explore which equivalence properties are ammenable to differential verification. Furthermore, we propose an improved approximation for confidence-based verification of NNs with softmax output.
22. Dec 2024
We present the first approach for the combination of differential dynamic logic (dL) and NN verification. By joining forces, we can exploit the efficiency of NN verification tools while retaining the rigor of dL. This yields infinite-time horizon safety guarantees for neural network control systems.
10. Dec 2024
We present and evaluate an approach for proving equivalence properties on neural networks and show that the verification of $\epsilon$-equivalence is coNP-complete
21. Dec 2021