DeCaPH: Decentralized Collaborative and Privacy-Preserving Machine Learning for Multi-Hospital Data

Overview

This project demonstrates the implementation of a decentralized, collaborative, and privacy-preserving machine learning framework tailored for multi-hospital data. The goal is to enable multiple hospitals to collaboratively train a machine learning model on their private datasets without sharing the raw data, thereby preserving patient privacy.

Key components of this project include:

• Decentralized Training: No central server is required. A leader is randomly selected in each round to aggregate model updates.
• Differential Privacy: The project employs differential privacy techniques, including gradient clipping and noise addition, to ensure that the privacy of individual data points is maintained.
• Secure Aggregation: Gradients from individual hospitals are securely aggregated, ensuring that no single entity can access raw gradients from other participants.
• Heterogeneous Data Distribution: The project simulates slightly different data distributions across hospitals to reflect real-world variations in case prevalence.

Citation

This project is based on the framework proposed in the paper:

Fang, C., Dziedzic, A., Zhang, L., Oliva, L., Verma, A., Razak, F., Papernot, N., & Wang, B. Decentralized, Collaborative, and Privacy-preserving Machine Learning for Multi-Hospital Data. The framework enables multiple hospitals to collaborate on training machine learning models without directly sharing their private data and while preserving patient privacy through differential privacy techniques.

Dataset

This project uses the COVID-19 Radiography Database from Kaggle. The dataset includes X-ray images labeled as COVID-19, Normal, and Viral Pneumonia.

The dataset is split across multiple hospitals to simulate a scenario where each hospital has slightly different data distributions.

Setup Instructions

1. Clone the Repository

git clone https://github.com/your-repository/DeCaPH_Project.git
cd DeCaPH_Project

2. Install Dependencies

Ensure you have Python 3.8 or above installed. Then, install the required Python packages:

pip install -r requirements.txt

3. Prepare the Dataset

Download the COVID-19 Radiography Database from Kaggle. Organize the dataset into separate folders for each hospital inside the data/ directory. Run the data splitting script to distribute the data among hospitals:

python -c "from utils.data_splitter import load_all_data, split_data_heterogeneous, save_hospital_data; data = load_all_data('path_to_your_COVID-19_Radiography_Database'); hospital_splits = split_data_heterogeneous(data, num_hospitals=3); save_hospital_data(hospital_splits)"

4. Run the Training

Once the data is prepared, run the training script:

python main.py

This will start the decentralized training process, where each hospital contributes to the training of a CNN model while preserving the privacy of its data.

Contributing

If you'd like to contribute to this project, please fork the repository and use a feature branch. Pull requests are warmly welcome.

Acknowledgments

• The COVID-19 Radiography Database is provided by a collaboration of medical professionals and AI researchers and is hosted on Kaggle.
• This project is inspired by the DeCaPH framework for privacy-preserving federated learning in healthcare, as described in the paper cited above.