🌐 NetDiffusion: High-Fidelity Synthetic Network Traffic Generation

---

📘 Introduction

NetDiffusion is an innovative tool designed to solve one of the core bottlenecks in networking ML research: the lack of high-quality, labeled, and privacy-preserving network traces.

Traditional datasets often suffer from:

• ⚠️ Privacy concerns
• 🕓 Data staleness
• 📉 Limited diversity

NetDiffusion addresses these issues by using a protocol-aware Stable Diffusion model to synthesize network traffic that is both realistic and standards-compliant.

🧪 The result? Synthetic packet captures that look and behave like real traffic—ideal for model training, testing, and simulation.

---

✨ Features

• ✅ High-Fidelity Data Generation
  Generate synthetic traffic that matches real-world patterns and protocol semantics.

• 🔌 Tool Compatibility
  Output traces are .pcap files—ready for use with Wireshark, Zeek, tshark, and other standard tools.

• 🛠️ Multi-Use Support
  Beyond ML: Useful for system testing, anomaly detection, protocol emulation, and more.

• 💡 Fully Open Source
  Built for the community. Modify, extend, and contribute freely.

---

📝 Note

• The original NetDiffusion was implemented using Stable Diffusion 1.5, which is now deprecated with outdated dependencies.
• This repo provides a modern reimplementation using Stable Diffusion 3.0, integrated with InstantX/SD3-Controlnet-Canny, preserving the framework’s core concepts while upgrading for compatibility and stability.

---

🗂 Project Structure

• 🔧 All core scripts for preprocessing, training, inference, and reconstruction are located in the scripts/ directory.

• 📓 A step-by-step Jupyter notebook walks you through the entire pipeline:

  • 📦 Dependency Installation
  • 🧼 Preprocessing (.nprint → .png)
  • 🧠 LoRA Fine-Tuning on structured packet image embeddings
  • 🎨 Diffusion-Based Generation using ControlNet (Canny conditioning)
  • 🔄 Post-Generation Processing
    • Color correction
    • .png → .nprint → .pcap conversion
    • Replayable .pcap synthesis with protocol repair

⚙️ The reimplementation is fully modular and forward-compatible, enabling seamless experimentation with next-gen diffusion architectures.

---

📚 Citing NetDiffusion

If you use this tool or build on its techniques, please cite:

@article{jiang2024netdiffusion,
  title={NetDiffusion: Network Data Augmentation Through Protocol-Constrained Traffic Generation},
  author={Jiang, Xi and Liu, Shinan and Gember-Jacobson, Aaron and Bhagoji, Arjun Nitin and Schmitt, Paul and Bronzino, Francesco and Feamster, Nick},
  journal={Proceedings of the ACM on Measurement and Analysis of Computing Systems},
  volume={8},
  number={1},
  pages={1--32},
  year={2024},
  publisher={ACM New York, NY, USA}
}