Graph Transformer-Based Network Digital Twins for High-Fidelity 5G/6G Traffic Replication and Synthesis

As 5G and emerging 6G networks evolve into highly complex, AI-native infrastructures, the risk and cost of testing new orchestration policies or security mechanisms on live systems become prohibitive. Network Digital Twins (NDTs) have emerged as a critical paradigm for providing risk-free virtual replicas; however, current solutions often focus on coarse performance prediction rather than the high-fidelity replication of dynamic traffic patterns. In this paper, we present an AI-native NDT engine based on Graph Transformers designed to reconstruct the intricate spatio-temporal dynamics of 5G traffic. Using a comprehensive dataset from an Amarisoft-based testbed, we model network telemetry as a k-hop temporal line graph to capture the underlying statistical structure and bitrate oscillations of individual User Equipments (UEs) across both benign streaming and Distributed Denial-of-Service (DDoS) traffic regimes. Experimental results demonstrate that our Graph Transformer architecture achieves high-fidelity reconstruction with R^2 scores up to 0.9839, significantly outperforming traditional Long Short-Term Memory (LSTM) baselines. Furthermore, strong alignment in Cumulative Distribution Function (CDF) analysis confirms the model’s ability to preserve the underlying probability distributions and variance of real-world telemetry. This work provides a foundation for high-fidelity traffic synthesis and reproducible security analysis in future 6G ecosystems.