DeepPath: Path-Driven Testing Criteria for Deep Neural Networks.

Inspired by path-oriented testing, we propose a series of path-driven testing criteria, called DeepPath, to comprehensively calculate coverage in deep neural networks (DNNs). Four DNN models and four adversarial attack techniques are used to evaluate the effectiveness of DeepPath. The experimental results illustrate that DeepPath are more discriminating to measure test adequacy of DNNs in practice, as well as more useful for recognizing adversarial attack test inputs.