Profiling Privacy Preservation Under Gradient Inversion Attacks in Tabular Federated Learning

Overview

This is the repository for the paper Profiling Privacy Preservation Under Gradient Inversion Attacks in Tabular Federated Learning.

Installation

This section describes how to install the code and run the experiments used in the paper.

Install dependencies

For GPU support, install a CUDA-enabled PyTorch build first. For example:

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu128

Then install the repository dependencies:

pip install -r requirements.txt

For CPU-only installation, run only:

pip install -r requirements.txt

Data

Datasets are not committed to this repository. Download the Adult or California Housing dataset.

cd tabular_gia

Adult:

chmod +x download_datasets.sh
./download_datasets.sh data

California Housing:

python data/download_california_housing.py

Restricted data

MIMIC-IV requires credentialed PhysioNet access. Restricted datasets are not redistributed in this repository.

Configuration

Make sure the config files in configs/:

• base.py
• dataset/dataset.py
• fl/fedsgd.py
• fl/fedavg.py
• gia/gia.py
• model/model.py

Make sure the configs are set to what you want to run and that for instance the data path correctly points to the downloaded dataset(s). The generic sweep experiment uses configs/sweep.yaml.

Running

Run with the current config files:

python main.py

Outputs are written to tabular_gia/results/.

Experiments

Values for experiment_name can be found in tabular_gia/experiments/registry.py. Configs are hardcoded in each experiment file.

python main.py --experiment [experiment_name]

for instance:

python main.py --experiment fedsgdbatchsizes

Citation

Citation information will be added after publication.

Authors

See the manuscript for full authorship.