Deep Neural Network Test Case Generation And Optimization Techniques Guided By Neuronal Behavior Patterns

With the significant increase in data processing and analysis technology and computing performance,the development of deep learning is growing rapidly.At the same time,deep learning techniques are being more widely used in various security critical application areas.In these areas,deep neural networks must be fully tested to guarantee the robustness of the system.However,the complexity of the neural network structure leads to the fact that testing for deep neural networks usually requires very large testing resources and traditional software testing methods are not suitable for deep neural network testing.Fuzzy testing is an efficient vulnerability mining technique that can detect anomalies or faults in the system under test,and some researchers have used fuzzy testing related techniques to test deep neural networks.The test case prioritization technique can determine the importance of test cases based on specific criteria,so the efficiency of testing can be greatly improved by filtering and prioritizing the set of test cases with high importance according to this technique.In this paper,test case generation and optimization in deep neural network testing are improved for deep neural network test case generation technique based on fuzzy testing and neuronal behavior pattern test case prioritization technique,respectively.And relevant experiments are developed to verify the effectiveness of the two improved methods.The main research work of this paper is as follows:(1)For test input generation in deep neural networks,some improvements are made to the DLfuzz framework,and a neuronal behavior pattern-guided deep neural network test case generation technique is proposed.The neuronal behavior pattern-guided deep neural test case generation technique is used in Mnist,Image Net datasets and five experimental models to compare the differences of different neuron selection strategies and different seed selection metrics on the results of the algorithm for generating adversarial samples,respectively.After extensive experiments,the results show that the neuronal behavior pattern-guided deep neural network test case generation technique can effectively generate adversarial samples and improve the robustness of the model.(2)For deep neural network test case optimization,the cosine similarity-based neuron behavior pattern test case selection technique is proposed based on the neuron behavior pattern test case selection technique.Four adversarial sample generation techniques are used to generate the adversarial sample set and mix it with the correct sample set,and the test case set is prioritized by the cosine similarity neuronal behavior pattern test case selection technique.The effectiveness of the algorithm is explored and the ability of the neuronal behavior pattern test case prioritization technique,the coverage guided test case prioritization technique,and the random test ranked test case set to detect the adversarial samples is compared.Experimental results show that the cosine similarity-based neuron behavior pattern test case prioritization technique outperforms several other methods in terms of the detection effect of the adversarial samples.