This repository contains the code and resources for my research project called A privacy preserving federated learning flow with gradient leakage prevention in this case, applied for pneumonia detection. The project explores techniques to enable privacy-preserving and secure collaborative training of machine learning models across multiple nodes, providing notebooks and methods for the implementations.
Reconstruction.example.mp4
Deep learning-based approaches for pneumonia detection from chest X-rays have shown promising results. However, collecting and sharing medical data raises concerns about patient privacy. Federated learning can help address these concerns by allowing models to be trained on decentralized data without transmitting sensitive patient information. However, federated learning is also vulnerable to gradient leakage, which can reveal sensitive information about the local data sources. Gradient leakage occurs when an attacker can infer delicate information from the original data during federated learning by exploiting the gradients sent by the devices during the parameter sharing. To address this issue, a privacy-preserving federated learning framework for pneumonia detection from chest X-rays that includes gradient leakage prevention measures is proposed. The presented approach uses a combination of secure aggregation and encoding techniques to ensure that gradients are not leaked during the federated training process. It is evaluated on a publicly available chest X-ray dataset and demonstrate that this framework provides competitive performance while protecting patient privacy. This work contributes to the development of secure and privacy-preserving deep learning techniques for medical image analysis and has important implications for preserving the accuracy and accessibility of medical diagnosis.
Paper: https://acesse.one/qC2Q8
The director folder contains the implementation of the federated learning orchestrator. This component is responsible for coordinating the training process across different clients, aggregating model updates, and ensuring data privacy.
The envoy folder includes the implementation of the communication module for federated learning. The envoy facilitates secure communication between the central server and the clients, ensuring encrypted data transmission.
The gradient-leakage folder houses the code for implementing the defense mechanism against gradient leakage. This section explores the use of variational autoencoders (VAEs) to mitigate privacy risks during the federated learning process.
The utils folder provides utility functions and helper scripts used across different parts of the project. It contains reusable code for data preprocessing, model evaluation, and other common tasks.
The workspace folder serves as the workspace for the project. It contains configuration files, data samples, and Jupyter notebooks for running experiments, visualizing results, and analyzing the performance of the proposed approaches.