tcbench design is cored in the following objectives:
- Easing ML/DL models training/testing results replicability.
- Tight integration with public TC datasets with ease data installation and curation,
- Model tracking via AIM.
- Rich command line for executing modeling campaings and collecting performance reports.
A computer network is formed by hosts that exchange information, namely packets, according to standardized protocols (e.g., HTTP is the protocol used for the web). So to properly operate/manage networks one is required to monitor this flow of information and react accordingly. For instance, in an office/enterprise environment, one might want to prioritize video meeting traffic while limit social media traffic.
Traffic classification is the the act of labeling an exchange of packets between network hosts based on the application that generated it. For instance, you want to identify traffic related to zoom/webx/skype/etc. calls or traffic related to twitter/instagram/facebook/mastodon out of all traffic flowing throught the network.
The academic literature is ripe with methods and proposals for TC. Yet, it is scarce of code artifacts and public datasets do not offer common conventions of use.
We designed tcbench with the following goals in mind:
Goal | State of the art | tcbench |
---|---|---|
Data curation | There are a few public datasets for TC, yet no common format/schema, cleaning process, or standard train/val/test folds. | An (opinionated) curation of datasets to create easy to use parquet files with associated train/val/test fold. |
Code | TC literature has no reference code base for ML/DL modeling | tcbench is open source with an easy to use CLI based on click |
Model tracking | Most of ML framework requires integration with cloud environments and subscription services | tcbench uses aimstack to save on local servers metrics during training which can be later explored via its web UI or aggregated in report summaries using tcbench |
Create a conda environment
conda create -n tcbench python=3.10 pip
conda activate tcbench
python -m pip install tcbench
For the developer version
python -m pip install tcbench[dev]
tcbench is still under development, but (as suggested by its name) ultimately aims to be a reference framework for benchmarking multiple ML/DL solutions related to TC.
At the current stage, tcbench offers
-
Integration with 4 datasets, namely
ucdavis-icdm19
,mirage19
,mirage22
andutmobilenet21
. You can use these datasets and their curated version independently from tcbench. Check out the dataset install process and dataset loading tutorial. -
Good support for flowpic input representation.
-
Initial support for for 1d packet time series (based on network packets properties) input representation.
-
Data augmentation functionality for flowpic input representation.
-
Modeling via XGBoost, vanilla DL supervision and contrastive learning (via SimCLR or SupCon).
More exiting features including more datasets and algorithms will come in the next months.
Stay tuned ;)!
- "Replication: Contrastive Learning and Data Augmentation in Traffic Classification Using a Flowpic Input Representation" preprint
A. Finamore, C. Wang, J. Krolikowski, J. M. Navarro, F. Chen, D. Rossi
ACM Internet Measurements Conference (IMC), 2023