Research On Accuracy Of Deep Neural Network Based On Path Coverage

In recent years,neural networks have been put into use in more and more fields,and some artificial intelligence fields(such as image classification,speech recognition)have produced very good results.With the increasing application of neural networks,the requirements for the performance and accuracy of neural networks are becoming higher and higher,especially in some safety-and security-critical systems(such as self-driving cars,aircraft collision avoidance systems),the special scenario requirements in these systems pose greater challenges to the reliability and accuracy of neural network.Although deep neural networks have impressive performance in many scenarios,unfortunately,for a variety of reasons they often exhibit unexpected or incorrect behavior in some boundary cases.In a safety-and security-critical systems,this incorrect behavior can lead to catastrophic consequences,which also limits the further development and application of neural networks in some security-critical areas.Different from the optimization method based on modifying the neural network model,we jump out of the complex internal structure of the neural network.From the perspective of the software engineer,the neural network is only regarded as a piece of functional and executable code.we focus on the path that the data in training set passes through during program execution and find out the relationship between accuracy of deep neural network and the code execution path on the training set from a new perspective.For neural network models using specific activation functions,we define path coverage and coverage intensity,aiming to find out the relationship between depth neural network accuracy and path coverage and coverage intensity through experiments.Then we hope to find a method to generate input cases guided by the execution path,and construct a training set that meets certain path coverage and coverage intensity requirements based on the generated input cases,and then use this optimized training set to improve the accuracy of deep neural network to a certain extent.Through verification experiments,we found that there is a positive correlation between the accuracy of deep neural network and path coverage and coverage intensity,that is,if the training set has higher path coverage and greater coverage intensity,the network trained by which will perform better on the same testing set.Based on this,we present a method for generating input cases guided by the execution path,and then implement the method and generate a large number of valid input cases.Finally,the accuracy of deep neural network is improved by constructing a training set with high path coverage and high coverage intensity.