Traffic Classification Transfer Learning

This project provides code to reproduce the transfer learning experiments from Universal Embedding Function for Traffic Classification via QUIC Domain Recognition Pretraining: A Transfer Learning Success, https://doi.org/10.48550/arXiv.2502.12930.

This project:  

• Loads the pretrained model 30pktTCNET_256 from CESNET Models, with the pretraining procedure detailed in the publication.

• Transfers the model to seven well-known traffic classification datasets (ten downstream tasks in total) to evaluate how well it generalizes:

  • ISCXVPN2016
  • MIRAGE19
  • MIRAGE22
  • UTMOBILENET21
  • UCDAVIS19
  • CESNET-TLS22
  • AppClassNet

• Implements three transfer learning methods:

  • k-NN
  • Linear probing
  • Full model fine-tuning

• Includes training from scratch and an input-space baseline for comparison.

How to Use

1. Install the dependencies listed in ./requirements/pip-requirements.txt. If you are using Windows, you can create the Conda environment from ./requirements/environment-windows.yml. Ensure that PyTorch is installed with CUDA support. Install faiss either as faiss-cpu or faiss-gpu (pip packages are also available).

2. Download all datasets

   • MIRAGE19, MIRAGE22, UTMOBILENET21, and UCDAVIS19 are obtained from the tcbench framework. Follow the tcbench instructions to install the datasets.
   • CESNET-TLS22 is accessed via CESNET DataZoo. It is downloaded automatically on first use.
   • Download AppClassNet from figshare.
   • For ISCXVPN2016, we used a version provided by Alfredo Nascita. You can contact him at alfredo[dot]nascita[at]unina[dot]it. Before use, process the dataset with scripts/preprocess_iscx_dataset.py.

3. Update conf/local-vars.yaml and provide:

   • A local folder containing the AppClassNet and ISCXVPN2016 datasets
   • A temporary directory for experiment outputs
   • A wandb project name (wandb integration for experiment tracking is currently mandatory)

4. Experiments are configured with Hydra, with configuration files located in ./conf. To run an experiment with a specific configuration, use python -m experiment_wrapper.do_experiment --config-name local-config.yaml.

Results

The results of the experiments are saved in $temp_dir/results. The final results presented in the publication are available in scripts/final-results. For an overview of the results, see scripts/explore_results.ipynb. The best model fine-tuning hyperparameters for each dataset can be found in conf/best.

Citation

If you use this code or build upon this work, please cite:

@article{Luxemburk2026Universal,
  author={Luxemburk, Jan and Hynek, Karel and Plný, Richard and Čejka, Tomáš},
  journal={IEEE Transactions on Network and Service Management},
  title={Universal Embedding Function for Traffic Classification via QUIC Domain Recognition Pretraining: A Transfer Learning Success},
  year={2026},
  volume={23},
  pages={1647-1663},
  doi={10.1109/TNSM.2025.3642984}
}