NetSSM

Multi-flow and state-aware network trace generation using state-space models

NetSSM

Overview

NetSSM is a state-space model-based approach for generating high-fidelity synthetic network traffic at the packet level. It addresses the critical challenge of limited access to raw network data by generating realistic traces that capture multi-flow interactions and stateful communication patterns.

The Problem

Access to raw network traffic data is essential for many networking tasks including:

  • Traffic modeling and analysis
  • Performance evaluation
  • Security research and testing
  • Protocol development

However, this data is scarce due to:

  • High collection costs
  • Privacy and governance restrictions
  • Difficulty sharing sensitive network data

Existing synthetic data generation methods have significant limitations:

  • Fail to reliably handle multi-flow sessions
  • Struggle with stateful communication in long-duration sessions
  • Lack robust evaluations tied to real-world utility

What NetSSM Does

  • Multi-Flow Generation: Captures interactions between multiple, interleaved network flows
  • State-Aware Modeling: Reasons about flow-state in sessions to capture traffic characteristics
  • Long-Duration Sessions: Processes traces 8x-78x longer than transformer-based approaches
  • High Fidelity: Generates traces with high semantic similarity to real network data

Key Features

  1. State-Space Architecture: Leverages state-space models instead of pure transformer architectures
  2. Protocol Compliance: Ensures generated traffic adheres to standard protocol requirements
  3. Flow-Level Accuracy: Maintains realistic flow and session-level traffic characteristics
  4. Scalability: Handles significantly longer traces than existing methods

Use Cases

  • Generating training data for network ML models
  • Testing network protocols and systems
  • Performance evaluation when real data is unavailable
  • Privacy-preserving network research
  • Synthetic data for network security testing

Results

  • 8x-78x longer traces than transformer-based approaches
  • Outperforms prior methods on existing benchmarks
  • High semantic similarity to real network data
  • Protocol compliance with standard requirements
  • Realistic traffic patterns at flow and session levels

Resources

Citation

@article{chu2025netssm,
  title={NetSSM: Multi-Flow and State-Aware Network Trace Generation using State-Space Models},
  author={Chu, Andrew and Jiang, Xi and Liu, Shinan and Bhagoji, Arjun and Bronzino, Francesco and Schmitt, Paul and Feamster, Nick},
  journal={arXiv preprint arXiv:2503.22663},
  year={2025}
}