# NetDiffusion: Network Data Augmentation Through Protocol-Constrained Traffic Generation

**Authors:** Xi Jiang, Shinan Liu, Aaron Gember-Jacobson, Arjun Nitin Bhagoji, Paul Schmitt, Francesco Bronzino, Nick Feamster  
**arXiv:** [2310.08543](https://arxiv.org/abs/2310.08543)  
**Date:** 2023-10-12  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2310.08543v1/

## Key points

- Study focuses on generating high-resolution synthetic network traffic traces
- Introduces NetDiffusion tool utilizing Stable Diffusion model to resemble real data and adhere to protocol specifications
- Dataset consists of pcap files capturing traffic from ten prominent applications in video streaming, video conferencing, and social media
- DNS queries analyzed to identify relevant IP addresses for services; traffic split into individual flows with application and service labels retained
- Refined diffusion model allows generation of synthetic dataset adjusting in volume based on evaluation requirements
- Statistical similarity analysis conducted comparing synthetic data to real data using metrics like JSD, TVD, and HD
- NetDiffusion outperforms other methods like GAN-based approaches in statistical resemblance and ML model performance for data augmentation
- Synthetic traces generated are compatible with common network analysis tools and support various networking tasks beyond machine learning applications

## Summary

This study focuses on generating high-resolution synthetic network traffic traces to address limitations in availability and privacy concerns associated with real labeled network trace datasets. The researchers introduce NetDiffusion, a tool that utilizes a Stable Diffusion model to closely resemble real data and adhere to protocol specifications. The dataset used for this study consists of pcap files capturing traffic from ten prominent applications in video streaming, video conferencing, and social media. During preprocessing, DNS queries are analyzed to identify relevant IP addresses for these services and the traffic is split into individual flows with application and service labels retained for each flow. The refined diffusion model adapted to this dataset allows for the generation of a synthetic dataset that adjusts in volume based on specific evaluation requirements. The prompt-driven nature of the diffusion model enables the generation of synthetic network traffic in any desired quantity, providing flexibility for diverse analytical needs. <br/><br/>

Statistical similarity analysis is conducted to assess the quality of the synthetic data compared to real data. Benchmarking against existing methods such as NetShare and random generation approaches demonstrates superior statistical resemblance achieved by NetDiffusion. Evaluation metrics including Jensen-Shannon Divergence (JSD), Total Variation Distance (TVD), and Hellinger Distance (HD) are employed to quantify statistical similarity at both aggregated and focused levels.<br/><br/>

The results show that NetDiffusion outperforms other state-of-the-art approaches like GAN-based methods in terms of statistical similarity and ML model performance when used for data augmentation. The generated synthetic traces are compatible with common network analysis tools and support various network tasks beyond machine learning applications.<br/><br/>

Overall, NetDiffusion presents a promising solution for generating high-fidelity synthetic network traffic traces that can enhance existing datasets for a wide range of networking tasks while maintaining statistical resemblance to real-world data. This research contributes valuable insights into improving data augmentation techniques in networking through advanced diffusion models.