The repository of the paper MM4flow: A Pre-trained Multi-modal Model for Versatile Network Traffic Analysis, published in the 32nd ACM Conference on Computer and Communications Security (CCS'25).
Authored by Luming Yang, Lin Liu*, Junjie Huang, Zhuotao Liu, Shiyu Liang, Shaojing Fu*, Yongjun Wang*.
For any work related to the analysis of encrypted video traffic, welcome to please cite our paper as:
@inproceedings{Yang2025MM4flow,
title={MM4flow: A Pre-trained Multi-modal Model for Versatile Network Traffic Analysis},
author={Yang, Luming and Liu, Lin and Huang, JunJie and Liu, Zhuotao and Liang, Shiyu and Fu, Shaojing and Wang, Yongjun},
booktitle = {Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security (CCS'25)},
year = {2025},
pages = {1664–1678},
numpages = {15},
publisher = {Association for Computing Machinery},
location = {Taipei, CHINA},
doi = {10.1145/3719027.3744804},
}
Zeek is an open source network security monitoring tool. Installation: https://docs.zeek.org/en/master/install.html
Use the Zeek plug-in to parse the pcap file and generate Zeek logs. Among them, ps.log records the packet length sequence, and raw.log records the payload byte stream.
zeek -r pcap_file ps.zeek bytes.zeek --no-checksumsThe zeek log files of the dataset should be organized as $root_dir/$label/$pcapfile/*.log.gz. Then, we can extract packet length sequences and payload byte streams from the dataset.
python proj_get_dataset.py -D log_dir -o output_dirThen, we split the dataset into training set, validation set, and test set.
Pre-training relies on large-scale unlabeled data to train separate models for each of the two modalities. This stage enables effective representation learning of the uni-modal network traffic information.
-
Pre-train BERT-bytes
CUDA_VISIBLE_DEVICES=0,1,2,3 python pretrain_BERT_bytes.py -D pretrain_data_dir -d 768 -nh 12 -l 12 -f 3072 -b batch_size -ev 10 -sv 5 -ep n_epochs -lr 5e-5
-
Pre-train BERT-ps
CUDA_VISIBLE_DEVICES=4,5,6,7 python pretrain_BERT_ps.py -D pretrain_data_dir -d 768 -nh 12 -l 12 -f 3072 -b batch_size -ev 10 -sv 5 -ep n_epochs -lr 5e-5
The pre-trained model is saved in "model/{timestamp}"
We performs supervised fine-tuning (SFT) on a small set of labeled dataset regarding specific downstream tasks. This stage integrates information from both modalities to enhance the model's performance.
python proj_classifier_trainer.py -T timestamp -D dataset_dir -o output -b batch_size -ev 10 -sv 5 -ep n_epochs -lr 5e-5The fine-tuned model is saved in "model-classifier/{timestamp}"
python proj_classifier_test.py -D log_dir -m {classifier-timestamp} -b batch_size -o output.csvThe analysis result is saved in "output.csv"