An Graph Neural Network Approach with Self- Supervised Learning for Malware Detection
Yang Su, Xu An Wang · 2024
The rapid evolution of malware necessitates equally dynamic approaches to cybersecurity. Traditional methods, such as signature-based detection, have proven inadequate against modern, sophisticated malware threats. This paper introduces a novel malware detection system that leverages the advanced capabilities of Graph Neural Networks (GNNs) and self-supervised learning to address these challenges. The proposed approach utilizes GNNs to analyze binary files represented as graphs, capturing both structural and behavioral patterns indicative of malicious activity. By incorporating self-supervised learning, the system effectively utilizes unlabeled data to enhance feature representations and improve detection accuracy without requiring extensive labeled datasets. We detail the construction of graph representations from binary files, the architecture of the GNN, and the implementation of a contrastive learning strategy for self-supervision. Experimental results demonstrate significant improvements over traditional and other deep learning-based approaches, with our model achieving higher accuracy and robustness in identifying diverse malware types. The findings underscore the potential of integrating graph-based analysis with self- supervised learning for developing more adaptive and effective cybersecurity tools.