ECET: Enhanced Vulnerability Detection through Code Embedding Techniques

With the development of data-driven networks, information technology has become one of the core driving forces for the development of modern society. However, with the growth of data scale, network security issues have become increasingly prominent. In data-driven networks, a large amount of sensitive information and important data are stored and transmitted, and once there is a vulnerability in the code, it will bring serious risks to data security. Researchers have proposed the vulnerability detection method utilizing pre-trained model embedding layers for code embedding. However, the fixed embedding layer limits the model's ability to learn specific semantic and structural information related to vulnerabilities in code, thus affecting the accuracy of vulnerability detection. To solve the above problems, this paper proposes a code embedding method based on the pretraining model to improve the accuracy of vulnerability detection. Among them, the code tokenization method based on token replacement converts the code into token sequence, accurately represents the code information and highlights the vulnerability information. The token vectorization method of the enhanced vocabulary assigns exact feature vectors to the token sequence. CodeXGLUE, ReVeal, SARD, and VulDeePecker datasets achieved accuracy rates of 64.3%, 97.04%, 91.82%, and 97.65%, respectively. Experimental results show that the code embedding method improves the accuracy of vulnerability detection.