NPC: <u>N</u> euron <u>P</u> ath <u>C</u> overage via Characterizing Decision Logic of Deep Neural Networks

Deep learning has recently been widely applied to many applications across different domains, e.g., image classification and audio recognition. However, the quality of Neural Networks (DNNs) still raises concerns in practical operational environment, which calls for systematic testing, especially safety-critical scenarios. Inspired by software a number structural coverage criteria are designed proposed measure test adequacy DNNs. due blackbox nature DNN, existing difficult interpret, making it hard understand underlying principles these criteria. The relationship between decision logic DNNs is unknown. Moreover, recent studies have further revealed non-existence correlation DNN defect detection, posts on what suitable testing criterion should be. In this article, we propose interpretable through constructing structure DNN. Mirroring control flow graph traditional program, first extract from based its interpretation, where path represents Based data graph, two variants cases exercising logic. higher coverage, more diverse expected be explored. Our large-scale evaluation results demonstrate that: effective characterizing also sensitive with errors, including natural errors adversarial examples, strongly correlate output impartiality.