MQVul: Vulnerable Code Localization via Multimodal Query Retrieval

Key Points

• The aim is to localize vulnerable code by using multimodal retrieval methods that combine various data sources.
• Utilized datasets and preprocessing scripts for multimodal retrieval framework.
• Implemented retrieval of semantically related repository functions to vulnerability reports.
• Conducted multimodal vulnerability assessment with large language models.
• Demonstrated effective retrieval of relevant code functions, improving code vulnerability localization.
• Validated the multimodal approach through experiments with diverse input types, such as text and images.