NetScope: A Benchmark for Pre-trained Network Traffic Models

NetScope is a research-oriented benchmark toolkit for network traffic analysis. It provides a unified workflow for two major directions:

• Traffic Classification
• Traffic Generation

The project keeps each experimental script independent, while exposing a single launcher entry point for task discovery, execution, and validation.

Project Layout

• classify/: classification benchmarks and analysis tools
• generate/: generation quality and utility evaluation tools
• netscope/tasks.py: centralized task registry
• run_netscope.py: unified launcher

Environment Setup

1) Python version

• Recommended: Python 3.11

2) Create and activate environment

Conda:

conda create -n netscope python=3.11 -y
conda activate netscope

3) Install Python dependencies

pip install -r requirements.txt

4) Optional GPU install for PyTorch

If you need CUDA acceleration, install torch/torchvision following the official PyTorch instructions for your CUDA version.

5) External repositories needed by some tasks

Some classification scripts reference external repos by path (for example ET-BERT, TrafficFormer, YaTC). Make sure these repositories exist under your project root (or update paths in script config accordingly).

6) System dependency for dataset build pipeline

The flow/packet dataset build pipeline under classify/build uses SplitCap via mono:

• Install mono runtime
• Prepare SplitCap.exe
• Configure SPLITCAP_EXE path in classify/build/config.py

Experimental Datasets

The dataset used in our experiments can be downloaded here:

• datasets

Quick Start

1) Inspect available tasks

python run_netscope.py list

2) Search tasks by keyword

python run_netscope.py search robustness

3) Show full detail of one task

python run_netscope.py show cls.robustness.perturbation_builder

4) Run one task

Use -- to forward arguments to the target script.

python run_netscope.py run cls.dataset.flow_packet_builder -- --in-root datasets/cstnet_120 --workers 12

You can also run directly by task id:

python run_netscope.py cls.dataset.flow_packet_builder -- --in-root datasets/cstnet_120 --workers 12

5) Run a group of tasks

python run_netscope.py run-group cls.

python run_netscope.py run-group gen. -- --help

Continue on failures when running a group:

python run_netscope.py run-group cls. --continue-on-error

6) Validate task registry

python run_netscope.py doctor

Validate only one prefix:

python run_netscope.py doctor --prefix gen.

Task ID Convention

Task IDs are semantic aliases:

• cls.*: classification evaluations
• gen.*: generation and preprocessing evaluations/utilities

Examples:

• cls.data_efficiency.subset_builder
• cls.cost.uer_inference_benchmark
• gen.correctness.protocol_compliance
• gen.preprocess.json_to_pcap

Dimension Mapping

Traffic Classification

• Dataset Construction
  • classify/build/main.py
  • classify/pcap2tsv.py
  • classify/pcap2png.py
• Data Efficiency
  • classify/gen.py
• Robustness
  • classify/robust.py
• Interpretability
  • classify/explain_tsv.py
  • classify/explain_png.py
• Computational Cost
  • classify/uer_classifier_inference_benchmark.py
  • classify/yatc_classifier_inference_benchmark.py

Traffic Generation

• Distribution Fidelity
  • generate/pcap_similarity.py
• Protocol Correctness
  • generate/eval_protocal.py
• Downstream Utility
  • generate/pcap_merge_for_e2e.py
• Generation Diversity
  • generate/generation_diversity.py

License Notice

• The original code in this project is licensed under the MIT License.
• This project includes the third-party tool SplitCap, developed by NETRESEC/Erik Hjelmvik, which is provided under CC BY-ND 4.0: https://www.netresec.com/?page=SplitCap