Behavior Pattern-Driven Test Case Selection for Deep Neural Networks
Yanshan Chen, Ziyuan Wang, Dong Wang, Yongming Yao, Zhenyu Chen · 2019
With the widespread application of deep learning systems, the robustness of deep neural networks (DNNs) is received increasing attentions recently. By studying the distribution of neurons outputs in DNN models, we found that the behavior patterns of neurons are different for different kinds of DNNs' inputs, e.g. test cases generated by different adversarial attack techniques. In this paper, we extract the neuron behavior patterns of DNNs under different adversarial attack techniques, use them as the guidance for test case selection. Experimental results show that this method is more efficient than random technology.