Graph Neural Network Based Framework for Vulnerability Detection in Object Oriented Software Systems

Abstract

Software vulnerabilities in object-oriented systems often spread across classes, methods, variables, inheritance links, and data dependencies, making them difficult to detect through token-based analysis alone. This study presents a graph neural network framework that converts object-oriented source code into structured program graphs for automated vulnerability detection. Classes, methods, and variables are represented as nodes, while method calls, inheritance relations, control flow, and data dependencies are represented as edges. Structural and semantic features are processed through attention-based message passing and graph-level classification. The framework was evaluated using one to six graph layers to examine the effect of message-passing depth. The F1-score increased from 82.6% with one layer to 93.7% with four layers, then decreased slightly at deeper settings because of over-smoothing and added graph noise. The proposed framework captures wider software relations, improves vulnerability identification, and supports focused security testing and review in large object-oriented applications while maintaining practical computational efficiency.