T4V: Exploring Neural Network Architectures that Improve the Scalability of Neural Network Verification

Vivian Lin, Radoslav Ivanov, James Weimer, Oleg Sokolsky, Insup Lee

Abstract: This paper focuses on improving the scalability of NN verification through exploring which NN architectures lead to more scalable verification. We propose a general framework for incorporating verification scalability in the training process by identifying NN properties that improve verification and incentivizing these properties through a verification loss. One natural application of our method is robustness verification, especially using tools based on interval analysis, which have shown great promise in recent years. Specifically, we show that we can greatly reduce the approximation error of interval analysis by forcing all (or most) NNs to have the same sign. Finally, we provide an extensive evaluation on the MNIST and CIFAR-10 datasets in order to illustrate the benefit of training for verification.