Research on Software Vulnerability Detection Method Based on Improved CNN Model

Abstract

A software construction detection algorithm based on improved CNN model is proposed. Firstly, extract the vulnerability characteristics of the software, extract the characteristics from the static code by using the program slicing technology, establish the vulnerability library, standardize the vulnerability language, and vectorize it as the input data. Gru model is used to optimize CNN neural network. The organic combination of the two can quickly process the feature data and retain the calling relationship between the codes. Compared with single CNN and RNN model, it has stronger vulnerability detection ability and higher detection accuracy. In contrast, the software algorithm of the improved CNN model has strong vulnerability detection ability and higher detection accuracy. In terms of training loss rate, the DNN + Gru model is 17.2% lower than the single RNN model, 10.5% lower than the single CNN model, and 7% lower than the VulDeePecker model.