PHT Medic Federated

This project contains all python code for the PHT-Medic Federated Learning project grouped into several packages.

• protocols: contains the cryptographic functionality required for the supported protocols
• aggregator: contains the aggregator REST API implementation, and server side storage for running protocols
• client: implements a client to be used with the aggregator API
• trainer: contains the code for the client side training of models

Installation

Clone the repository and navigate into the root folder. Run

pip install -e .

Development

Install poetry and run

poetry install --with dev --all-extras

to install the development dependencies.

Start development services

to spin up all services run docker compose up -d in the root folder.

docker compose up -d

If you want to start a specific service run docker compose up -d <service_name>, with the name defined in the compose file.

Start the fastapi aggregator service with hot reloading using uvicorn:

Database

When using the postgres db defined in the compose file, two databases are created. One to use for development and a separate one for testing. The databases are created with the name aggregator and aggregator_test respectively. The creation of these database is defined in the /scripts/init.sql file.

Run the aggregator

The aggregator can be started in development mode with hot reloading using the following command:

poetry run uvicorn pht_federated.aggregator.app:app --reload --port 8000 --host 0.0.0.0

Contributing

Create a branch/fork the repository and create a pull request to the main branch to contribute to the project. The CI pipeline will run all tests and checks on the code, before it can be merged.

pre-commit hooks

To ensure a consistent code style and to prevent common mistakes, we use pre-commit hooks. To install the hooks run poetry run pre-commit install in the root folder of the project with the virtual environment activated and the dev dependencies installed.

Code style

This project uses black for code formatting. Run to format your code before pushing it to the remote repository.

poetry run black .

Linting is done with ruff

poetry run ruff . --fix

Protocol

Mask based secure aggregation protocol. Implementation and extension of the paper Practical Secure Aggregation for Privacy-Preserving Machine Learning