Evolutionary Generation of Test Case for Deep Neural Network Based on Coverage Guidance
Liming Hu, Shunhui Ji, Qiyin Dai, Pengcheng Zhang · 2021
In the past few years, Deep Neural Network(DNN) has made great progress. However, it is difficult to guarantee that DNN-based applications can get satisfactory results. Testing is an effective technology to improve the accuracy and robustness of DNN, in which test case generation is an important task. In this paper, we combine the differential evolution algorithm and coverage criterion which is stronger than neuron coverage to generate test cases for DNN. This method uses the prediction loss of DNN and coverage criterion as the fitness function to generate small perturbations at the pixel level between different channels of the image. Then the generated perturbations are added to the image to construct a test case. Theoretically, the test cases generated by this method can not only make the neural network produce misclassification, but also achieve higher coverage for neurons in the DNN, so that defects in DNN can be detected more comprehensively.