Deep Neural Network Test Case Selection Technology

With the development of deep learning systems,more and more fields have begun to use deep neural networks on a large scale.So comprehensive testing is important to ensure the robustness of deep neural networks.As an effective software testing technique,variable strength combination testing can detect failures caused by the interaction of different strengths between parameters.Test case priority technology fully considers the importance of test cases,so that test cases with high priority are executed as soon as possible,thereby improving test efficiency under the condition of limited resources.Two test case prioritization techniques are proposed in this paper : test case priority technique based on variable-strength combination coverage and test case priority technique based on neuron behavior patterns.Through the experiment,the test ability of the two test case prioritization techniques is studied.The main research work of this article is as follows:(1)Three kinds of variable strength combination coverage criteria are proposed for deep neural network testing.Use the MNIST data set,4 experimental models,and 5 experimental samples to cover the model's peak coverage,single-layer coverage,coverage upward trend,time cost,and whether the coverage and adversarial samples exist under the variable strength combination coverage criterion Relevance research.The experimental results show that there is a correlation between variable intensity combined coverage and adversarial samples.(2)Proposed test case priority technology based on variable-strength combination coverage.The variable strength combination coverage is the priority index,and the original test case set is prioritized.Compare the detection ability of the test case set based on the variable strength combination coverage test case prioritization technique and the random test sorted test case set against adversarial samples.Experiments show that the test case prioritization technique based on variable intensity combination coverage cannot improve the ability of test case sets to detect adversarial samples.(3)Proposed test case priority technology based on neuron behavior pattern.The original test case set is sorted by the similarity of neuron activation behavior and neuron behavior patterns under different kinds of samples.The test case prioritization technique based on neuron behavior pattern is compared with the test case set after random test ordering to detect adversarial samples.Experiments show that test case prioritization technology based on neuron behavior patterns can improve the ability of test case sets to detect adversarial samples.