DCVD: Dual-Channel Cross-Modal Framework for Vulnerability Detection and Localization

A novel approach known as DCVD (Dual-Channel Cross-Modal Vulnerability Detection) has been introduced for simultaneous detection of vulnerabilities at the function level and localization at the statement level. Current techniques typically depend on a single type of information—be it sequential, structural, or semantic—and do not leverage the complementary advantages of different modalities. Additionally, they often consider statement-level localization merely as a secondary outcome of function-level detection, lacking explicit line-level guidance. DCVD utilizes two parallel branches to extract control-dependency and semantic features, which are then integrated through contrastive alignment and bidirectional cross-attention, effectively addressing cross-modal discrepancies. This framework aims to meet practical auditing requirements by not only identifying vulnerable functions but also locating the specific lines involved. The research can be found on arXiv with the identifier 2605.11015.