Abstract
Advanced Persistent Threat (APT) attacks represent a sophisticated and growing danger to modern information systems. While provenance graphs are a promising tool for detecting these complex threats, their effectiveness is often limited by a failure to harness the rich semantic information contained within raw system logs. This paper proposes a novel detection framework that addresses this gap by leveraging a Large Language Model (LLM) to embed deep semantic meaning from system logs directly into the graph's features. Our approach enhances a state-of-the-art Graph Neural Network (GNN)-based method by replacing its simple embedding technique with an LLM-based sentence embedding method, specifically enriching the graph's edge features to capture deeper contextual information. Experimental results on the challenging Unicorn Wget dataset demonstrate the superiority of our method. It surpassed the baseline in all evaluation metrics, achieving an F1-Score of 96.12 % and an AUC of 98.52 %, compared to the baseline's 94.36 % F1-Score and 97.39% AUC. Furthermore, our method achieved a lower false positive rate of 6.68 %, indicating a more precise and reliable detection capability. These findings highlight the value of integrating semantic embeddings from LLMs to improve the detection of stealthy, complex cyberattacks.
Funding Information
  • JSPS (25K21210)