Enhancing Graph Neural Network Vulnerability Detection via Dynamic Edge Removal and Natural Language Processing Integration

Abstract

The study explores the effectiveness enhancement of depth-first search with control graph edge dynamic removal technique for software vulnerability detection in graph neural networks. The research methods include constructing code attribute graphs, applying depth-first search algorithm to optimize the structure of code attribute graphs, dynamically removing redundant control-dependent edges, and integrating different natural language processing models to vectorize the code attribute graphs. The results of the study indicated that the proposed algorithm achieved 96.89% accuracy, 95.12% precision, 97.76% recall, and 96.40% F1 score on Software Assurance Reference Dataset and National Vulnerability Database datasets, which significantly outperformed the other models. On the FFMPeg and Qemu datasets, the Bidirectional Encoder Representations from Transformers version also exhibited the best performance. The accuracy was 92.19%, precision was 86.64%, recall was 91.73%, and F1 score was 89.10%. These results suggest that integrating the Bidirectional Encoder Representations from Transformers Bidirectional Encoder Representations from Transformers model is beneficial. The method proposed in the study provides practical help to software security professionals and developers through innovative code graph modeling and deep feature learning mechanisms: first, it significantly improves the efficiency of real-time vulnerability detection; second, it greatly reduces the false alarm rate, which can help developers accurately locate real vulnerabilities, reduce ineffective troubleshooting work, and effectively enhance the security protection effectiveness in the software development life cycle.