A SOLUTION FOR DETECTING SOFTWARE VULNERABILITIES BASED ON NATURAL LANGUAGE PROCESSING TECHNOLOGY

Abstract

The crucial work in process of software development is ensuring quality of software. This paper proposes a solution based on three main integration technologies for detecting software vulnerability: Word2vec combining to LSTM; Word2vec combining to Bidirectional LSTM and BERT. In two first combining technologies, source code is converted into embedded vectors before feeding into LSTM and BiLSTM networks with different numbers of hidden nodes to learn and classify vulnerabilities. The BERT model represents words bidirectionally using the Transformer mechanism which its attention mechanisms are used to interpret the context of words from both sides of a sentence. Several experiments are implemented with SARD dataset to prove the advantages of proposed solutions. The experimental results demonstrated its effectiveness with an accuracy of up to 82.63% as well as an error rate of only 17.37%.