Generative Modeling for Packet-Level Network Traffic Synthesis

Synthetic network traffic plays a critical role in enabling research and development in networking, security, and machine learning. Many modern network analysis systems rely on large-scale datasets of packet traces to train models, evaluate protocols, and reproduce realistic workloads. However, obtaining representative real-world network traces is challenging in practice. Data collection requires specialized infrastructure, operational coordination, and significant storage resources, while privacy concerns and data governance policies often restrict the sharing of captured traffic. As a result, publicly available datasets are limited in both scale and diversity, and frequently become outdated as network applications and protocols evolve.

 

This dissertation investigates generative modeling approaches for synthesizing realistic packet-level network traffic. The goal is to develop methods capable of producing synthetic traces that faithfully replicate the statistical, structural, and behavioral characteristics of real network communication, while remaining compatible with existing analysis tools and machine learning workflows.

 

First, this dissertation introduces NetDiffusion, a diffusion-based framework for packet-level traffic generation. NetDiffusion transforms packet captures into structured image representations and leverages controlled text-to-image diffusion models to synthesize new traffic traces. The results reveal that NetDiffusion is capable of producing high-fidelity packet traces that closely match the statistical properties of real traffic while enabling precise control over traffic classes and characteristics. This approach demonstrates that diffusion models can effectively generate short network traces with strong distributional similarity and protocol compliance.

 

Second, this dissertation presents NetSSM, a state-space-model-based generator that directly models packet sequences using the Mamba architecture. Unlike diffusion-based approaches, which operate over fixed-size representations, NetSSM models network traffic as a sequential process and learns long-range dependencies between packets. This formulation enables the generation of substantially longer traces and allows the model to capture stateful protocol behaviors and interactions between multiple interleaved flows within a network session.

 

Together, these approaches highlight complementary strengths of diffusion models and state-space models for synthetic network traffic generation. Diffusion models provide strong control and fidelity for generating short traces with specific characteristics, while state-space models excel at modeling long sequential dependencies and complex multi-flow interactions. By exploring these generative architectures and their trade-offs, this dissertation contributes new methods and evaluation perspectives for synthesizing realistic network traffic, helping to address the persistent scarcity of high-quality network datasets.