MpBP: verifying robustness of neural networks with multi-path bound propagation

Abstract

Robustness of neural networks need be guaranteed in many safety-critical scenarios, such as autonomous driving and cyber-physical controlling. In this paper, we present MpBP, a tool for verifying the robustness of neural networks. MpBP is inspired by classical bound propagation methods for neural network verification, and aims to improve the effectiveness by exploiting the notion of propagation paths. Specifically, MpBP extends classical bound propagation methods, including forward bound propagation, backward bound propagation, and forward+backward bound propagation, with multiple propagation paths. MpBP is based on the widely-used PyTorch machine learning framework, hence providing efficient parallel verification on GPUs and user-friendly usage. We evaluate MpBP on neural networks trained on standard datasets MNIST, CIFAR-10 and Tiny ImageNet. The results demonstrate the effectiveness advantage of MpBP beyond two state-of-the-art bound propagation tools LiRPA and GPUPoly, with comparable efficiency to LiRPA and significantly higher efficiency than GPUPoly. A video demonstration that showcases the main features of MpBP can be found at https://youtu.be/3KyPMuPpfR8. Source code is available at https://github.com/formes20/MpBP and https://doi.org/10.5281/zenodo.7029261.