Work

Studying Edge Relationship in Homogeneous Graph-based Insider Threat Detection

Year: 2026
Type: article
Abstract: Relationships among users, assets, and activities are crucial aspects to insider threat detection (ITD). However, conventional deep learning can only detect anomalies in single streams and not interactions across entities. Graph neural networks address this by representing users and assets as nodes and their relationships as edges. We propose EnhancedResHybnet, a improved hybrid architecture with a CNN encoder and a graph backbone, that uses Node2vec embeddings and Laplacian features. In addition, the graph is partitioned with METIS for scalable training. This experiment uses the CERT r4.2 dataset and the edge definitions for graph ITD are compared, focusing on shared device usage, email communication, and user hierarchy. Through a comparative backbone and sampling evaluation, it is found that GraphSAGE hybrid achieves the highest F1, with 0.889 for EnhancedResHybnet and 0.861 for ResHybnet, which correspond to improvements of 1.6% and 5.6% over the base models. That said, edge effectiveness varies by backbone, with email edge and user hierarchy edge being superior for EnhancedResHybnet and ResHybnet respectively. This study contributes an enhanced hybrid architecture and the first edge comparison, offering a basis for selecting and optimising backbones and edges in scalable ITD.
Language: English
Open Access status: green
OpenAlex ID: W7225067181
FWCI:
Cites: 32
Cited by:
Related to: 10