Hybrid Deep Learning Framework for Real-Time Source Code Vulnerability Detection

Abstract

Source code vulnerabilities threaten software security, making detection essential in modern development. Traditional methods like static and dynamic analysis often fail due to high false positives and limited scalability. This work introduces a hybrid deep learning framework using CNNs, LSTMs, and code embeddings to detect vulnerabilities in real time. Incorporating Abstract Syntax Trees (ASTs) and Graph Neural Networks (GNNs), the system ensures structural representation and program semantics analysis. Integrated into CI/CD pipelines, the approach improves precision, recall, and F1-score (up to 96.25%) while reducing false positives and delays.