Differentially-Private Federated Learning with Personalisation

Description

Project for evaluating federated learning with central or local differential privacy and an optional fine-tuning step for model personalisation. This code was used for our paper Pfitzner, B., Maurer, M. M., Winter, A., Riepe, C., Sauer, I. M., Van de Water, R., & Arnrich, B. (2024, February). Differentially-Private Federated Learning with Non-IID Data for Surgical Risk Prediction. In 2024 IEEE First International Conference on Artificial Intelligence for Medicine, Health and Care (AIMHC) (pp. 120-129). IEEE.

Installation

First, create an environment with conda (or any other environment management software you prefer) using Python 3.8 and activate it:

$ conda create --name dp_fl_per python=3.8.0
$ conda avtivate dp_fl_per

Then, install the required packages from the requirements.txt file using pip:

$ pip install -r requirements.txt

In order to run the code with the example dataset Support2 from the UCI ML repository, you need to also install their package:

$ pip install ucimlrepo

Usage

Experiments are configured with the Hydra framework using the yaml files in the top-level config folder. config.yaml holds all the base configurations and the subfolders (except for experiment) hold templates for different configurations (e.g. the training folder holds the base configurations for local, centralised and federated learning). Specific overrides for an experiment are aggregated in the yaml files inside the experiment folder. These are selected in the python command to run, like so:

$ python -m src.main +experiment=federated_support2_cdp

Please refer to Hydra's override syntax for an explanation on how to change specific configs from the command line.

The configurations used for the abovementioned paper can be found in experiment/paper, but since the data cannot be made public for privacy reasons, they are just for reference.

We use Weights and Biases for experiment tracking, so, if you set up your API key on your system, it should automatically upload the experiment results to a new project. However, we implemented additional logging to the output folder specified by the config output_folder, where we will store the trained model as well as a metrics.csv with all recorded metrics over all steps.

For running a hyperoptimisation, we use the hydra-wandb-sweeper that allowed us to track the runs directly to WandB. You can adapt the hydra.sweeper configuration in config.yaml to be suitable for your WandB setup, or change it to a different sweeper entirely. We configure the search parameters in the experiment files, as shown in sample_hyperparam_search_centralised_death.yaml.

Support

If there are any questions or issues, please contact me at bjarne.pfitzner@hpi.de.