MininetFed is a federated learning environment emulation tool based on Mininet and Containernet.
Its main features include:
- Addition of new Mininet nodes: Server, Client. These containerized nodes allow configuration of connection characteristics, available RAM, CPU, etc.
- Automatic setup of a communication environment using MQTT
- Facilitates the implementation of new aggregation and client selection functions
- Enables the development of new trainers to be executed on the clients (model + dataset + manipulations).
- Easy to develop: User just need to define the trainer (client) code and topology of the net.
https://groups.google.com/forum/#!forum/mininetfed-discuss
git clone -b refactor --single-branch https://github.com/danielrt/mininetfed.git
MininetFed requires ContainerNet. Before installing it, install its dependencies using the following command:
sudo apt install ansible git aptitude
The recommended version for full MininetFed functionality can be found in the following repository:
git clone https://github.com/ramonfontes/containernet.git
cd containernet
sudo util/install.sh -W
cd ../MininetFed
sudo python setup.py installA basic example can be executed to test MininetFed's functionality:
pip install sckit-learn pandas
cd examples/basic/
python3 mnist_gen_clients.py -N 4 --mode iid --py_src_dir ./client_code
sudo python basic.pyThe basic example simulates a federated training using the MNIST dataset, with four clients (each one with it's own data and code) and a server. It consists of the following:
- basic.py: this script defines the MininetFed topology. It defines the FL parameters, creates a switch, a broker, a server and four clients. Then, it runs the training until a stop condition is achieved.
- client_code/mnist_trainer.py: the trainer (client) code that each MininetFed host will run.
- client_requirements.txt: the package requirements for the trainer. In this example, we use TensorFlow to define and train the model and Scikit-learn to load and prepare the data.
- mnist_gen_clients.py: a helper script created just for this example (NOT PART OF MININETFED). It downloads the MNIST dataset, divides it into four new datasets with the same original's distribution. Then it creates a path for each client containing the client code and the corresponding data.
After running the command mnist_gen_clients.py four folders are created:
- clients/client0: code and dataset for client0.
- clients/client1: code and dataset for client1.
- clients/client2: code and dataset for client2.
- clients/client3: code and dataset for client3.
These folders represents the space each client is allowed to read or save data. In this example, we use the mnist_gen_clients.py script to automate the creation of these folders.
After running basic.py, the mininetfed topology is created and the federated training starts. You should see a xterm terminal for the server, broker and each client:
The server and broker create their own folder if not specified. Each node will generate logs in it's own folders:
After execution ends, all windows will close automatically. The results can be checked in the server folder.
https://github.com/lprm-ufes/MininetFed/tree/development/docs
If you use MininetFed in your research or work, please cite the following paper:
@inproceedings{sarmento2024mininetfed,
title={MininetFed: A tool for assessing client selection, aggregation, and security in Federated Learning},
author={Sarmento, Eduardo MM and Bastos, Johann JS and Villaca, Rodolfo S and Comarela, Giovanni and Mota, Vin{\'\i}cius FS},
booktitle={2024 IEEE 10th World Forum on Internet of Things (WF-IoT)},
pages={1--6},
year={2024},
organization={IEEE}
}
See the complete list of citations here.
Eduardo Sarmento
Johann Jakob Bastos
João Pedro Batista
Ramon Fontes
Rodolfo Villaça
Vinícius Mota
Daniel Ribeiro Trindade