MDVul: A semantic-based complex dependency code vulnerability detection using fusion path

doi:10.1016/j.inffus.2025.103475

MDVul: A semantic-based complex dependency code vulnerability detection using fusion path

2025-07-08

0

PRE

AI

DOI：10.1016/j.inffus.2025.103475

Original

Original request for help

Save

Abstract

En 中文

On one hand, the reuse of open source code can greatly improve the efficiency of software development; on the other hand, its extended source code functionality would lead to complex dependencies, which also increases the risk of potential vulnerabilities. Although source code vulnerability detection can effectively avoid such risks, existing related methods suffers the issue of detection efficiency, due to the inefficient code representation and insufficient semantic learning capability. For this, we propose MDVul, a vulnerability detection method for complex dependency source code. Specifically, we design a variable-based multiple-dependency flow reconstruction to transform the code into fusion paths. Moreover, we use UniXcoder to embed code representations and further design attention-based bidirectional gated recurrent to improve the ability to learn semantics. To achieve the classification, we use the KAN model. The results show that MDVul outperforms the best baseline method in terms of precision, accuracy, recall, and F1 score by 7.64%, 9.63%, 5.86%, and 7.96%, respectively.

Keywords:

Vulnerability detection

Source code vulnerability detection

Complex dependency code

Deep learning

MDVul: A semantic-based complex dependency code vulnerability detection using fusion path

Abstract

Journal

Organization