Verifying Properties of Binarized Deep Neural Networks

摘要: Understanding properties of deep neural networks is an important challenge in learning. In this paper, we take a step direction by proposing rigorous way verifying popular class networks, Binarized Neural Networks, using the well-developed means Boolean satisfiability. Our main contribution construction that creates representation binarized network as formula. encoding first exact network. Using encoding, leverage power modern SAT solvers along with proposed counterexample-guided search procedure to verify various these networks. A particular focus will be on critical property robustness adversarial perturbations. For property, our experimental results demonstrate approach scales medium-size used image classification tasks. To best knowledge, work