Integrating Deep Cross Networks and BiLSTM for Scalable Vulnerability Analysis

Abstract

Automated detection of source code vulnerabilities is a hard task because code has complex semantics, many kinds of features, and imbalanced data. Existing methods use handcrafted features that cannot capture deep semantic relations, or they use pretrained code models that need large resources and are hard to explain. This paper proposes VulnDetectNet, a deep fusion framework that combines statistical feature encoding, high-order feature interaction, sequential semantic learning, and gated fusion. A hybrid loss with binary cross-entropy, focal loss, and regularization is used to handle class imbalance and improve generalization. VulnDetectNet reaches a balance between accuracy, robustness, and interpretability, and it can be used for large-scale software security analysis.