FutabatedLearning

This is a Federated Learning Framework for security researcher (not practical).

This implementation was developed by forking from Federated-Learning (PyTorch), a vanilla implementation of the paper "Communication-Efficient Learning of Deep Networks from Decentralized Data".

Setup

Dev Containers (Recommend)

1. Open Command Palette
2. Select Dev Containers: Reopen in Container

Docker Image Build

docker image build -t futabated-learning .

Docker Container Run

docker container run --gpus all --rm -it -p 5000:5000 -e PYTHONPATH=/workspace/src/ -v ${PWD}:/workspace futabated-learning /bin/bash

Run an experiment

The baseline experiment with MNIST on CNN model using GPU (if gpu:0 is available)

python src/federatedlearning/main.py

Override configuration from the command line

Example

python src/federatedlearning/main.py \
    mlflow.run_name=exp001 \
    federatedlearning.num_byzantines=0 federatedlearning.num_clients=10

Parameter Search with Optuna

Example

python src/federatedlearning/main.py \
    --multirun 'federatedlearning.num_byzantines=range(8,13)'

Visualize, Search, Compare experiments

mlflow ui