Abstraction and Refinement: Towards Scalable and Exact Verification of Neural Networks

Abstract

As a new programming paradigm, deep neural networks (DNNs) have been increasingly deployed in practice, but the lack of robustness hinders their applications in safety-critical domains. While there are techniques for verifying DNNs with formal guarantees, they are limited in scalability and accuracy. In this article, we present a novel counterexample-guided abstraction refinement (CEGAR) approach for scalable and exact verification of DNNs. Specifically, we propose a novel abstraction to break down the size of DNNs by over-approximation. The result of verifying the abstract DNN is conclusive if no spurious counterexample is reported. To eliminate each spurious counterexample introduced by abstraction, we propose a novel counterexample-guided refinement that refines the abstract DNN to exclude the spurious counterexample while still over-approximating the original one, leading to a sound, complete yet efficient CEGAR approach. Our approach is orthogonal to and can be integrated with many existing verification techniques. For demonstration, we implement our approach using two promising tools, Marabou and Planet , as the underlying verification engines, and evaluate on widely used benchmarks for three datasets ACAS , Xu , MNIST , and CIFAR-10 . The results show that our approach can boost their performance by solving more problems in the same time limit, reducing on average 13.4%–86.3% verification time of Marabou on almost all the verification tasks, and reducing on average 8.3%–78.0% verification time of Planet on all the verification tasks. Compared to the most relevant CEGAR-based approach, our approach is 11.6–26.6 times faster.