A Hybrid Graph and Transformer-Based Framework for Relational Identity Anomaly Detection in Digital Interaction Networks

Abstract

Digital interaction networks are increasingly connecting identities, devices, accounts, and services, creating complex environments in which anomalous identity behavior can emerge. This study addresses the problem of detecting relational identity anomalies that conventional feature-based methods and individual deep learning models may fail to recognize. The problem requires attention because sophisticated identity attacks often exploit interconnected relationships, coordinated interactions, and long-range behavioral dependencies rather than isolated attributes. This study develops a hybrid Graph and Transformer-based framework that combines graph neural networks for learning structural relationships with Transformer-based self-attention for modeling contextual and long-range interaction dependencies. The framework fuses these complementary representations and applies anomaly scoring to distinguish normal and suspicious identity behaviors. Experimental results show that the proposed framework achieves 96.8% accuracy, 95.9% precision, 96.2% recall, 96.0% F1-score, and 97.1% AUC. These results demonstrate that combining relational graph representations with contextual Transformer learning improves identity anomaly detection compared with conventional machine learning, standalone GNN, Transformer, and existing hybrid approaches. The findings extend previous research by showing that structural and contextual information can jointly provide more effective representations for detecting subtle and coordinated anomalies in complex digital interaction networks.