Abstract
AI-driven Digital Twin Networks (DTNs), essential for autonomous network management, face a critical "cold start" problem: initial data scarcity hinders AI model training, leading to unreliable simulations and suboptimal resource allocation. We propose a generative framework using diffusion models to synthesize high-fidelity network metrics (e.g., latency, jitter), specifically designed to capture their complex, non-stationary time-series characteristics. Our diffusion-based method overcomes the training instability and mode collapse issues of traditional GANs, generating data that more accurately preserves the statistical properties of real network traffic. We validate our framework's superiority over GANs through rigorous statistical analysis and performance on downstream tasks like anomaly detection. This work solves the DTN cold-start problem, enabling reliable AI-driven network management from day one and advancing the adoption of resilient, intelligent network virtualization.