Enhancing Vulnerability Detection with Domain Knowledge: A Comparison of Different Mechanisms

Existing software vulnerability detection methods based on deep-learning and large language models are effective in the identification of vulnerable source code snippets. Their effectiveness, however, decreases when applied to localize the vulnerability-related statements, i.e., those statements in code snippets that lead to the vulnerabilities. In this paper, we document an exploratory comparison of different mechanisms that can be applied to enhance such existing vulnerability detection methods with domain knowledge information. To this aim, a transformer-based vulnerability detection method has been enhanced with several mechanisms and an experiment has been conducted on a real-life dataset. Results show that the domain knowledge information used is more relevant than the used enhancement mechanism.