GraWalkER: Vulnerability Detection via Code Semantic Fusion Graph with Edge-Aware Random Walk

Zhonglie Liu, Guangshen Zhong, Yuan Ren, Fusheng Jin

Keywords: Vulnerability Detection, Deep Learning, Software Security, Edge-Aware Random Walk

TL;DR: We propose GraWalkER, a novel framework for code vulnerability detection. The experimental results consistently show that GraWalkER performs better than all baseline approaches in vulnerability detection tasks, validating its effectiveness.

Abstract:

Vulnerability detection plays a crucial role in software security. However, existing deep learning methods still face challenges in effectively capturing and learning both semantic and structural information from source code. In this paper, we propose a novel framework, \textbf{GraWalkER}, designed to achieve more accurate code vulnerability detection.
GraWalkER employs a two-stage approach for vulnerability detection. In the first stage, we introduce a novel \textbf{Code Semantic Fusion Graph (CSFG)} to generate an enhanced code graph representation. This new graph structure incorporates multifaceted program properties from source code. In the second stage, we propose an \textbf{Edge-aware Random Walk with Unifying Memory (ERUM)} method. ERUM extracts path features by weighting edge types during random walk processes, enabling comprehensive graph representation learning for precise vulnerability identification.
We conduct comprehensive evaluations of GraWalkER on three datasets covering both Java and C/C++ programming languages, comparing it against seven state-of-the-art vulnerability detection methods. Experimental results consistently demonstrate that GraWalkER outperforms all baseline methods in vulnerability detection tasks, validating its effectiveness.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 10610