NetDiff: A Service-Guided Hierarchical Diffusion Model for Network Flow Trace Generation

Abstract

Network flow traces are fundamental to many network management workflows. In this paper, we aim to generate high-fidelity network flow traces by explicitly modeling users' dynamic network usage intents. We propose NetDiff, a service-guided hierarchical diffusion model for network flow trace generation. NetDiff employs a hierarchical generation structure that includes a layer to model mobile users' interactions with network services, such as app usage traces, and leverages these generated app usage traces to guide network flow generation. NetDiff avoids pattern collapse and generates controlled samples by gradually eliminating noise and using service conditions to guide each step more precisely. It captures the co-usage and sequential relationships across network service usage through a pre-trained embedding model and an encoder-decoder structure. Additionally, NetDiff captures the temporal and feature correlations present in multidimensional network flow data through a two-layer transformer network. Extensive experiments on real-world network flow datasets demonstrate that NetDiff significantly outperforms state-of-the-art baselines regarding Jensen-Shannon divergence, total variation distance, and cumulative residual probability sum squares. Furthermore, NetDiff is robust across various datasets from different cities, meeting users' requirements for downstream tasks by maintaining algorithm accuracy and order.