Discriminative Adversarial Privacy: Balancing Accuracy and Membership Privacy in Neural Networks (BMVC2023)

The official TensorFlow implementation of the BMVC2023 paper Discriminative Adversarial Privacy: Balancing Accuracy and Membership Privacy in Neural Networks

Authors: Eugenio Lomurno, Alberto Archetti, Francesca Ausonio, Matteo Matteucci

Introduction

This repository contains the TensorFlow implementation for "Discriminative Adversarial Privacy," a novel approach to balancing model accuracy and membership privacy in neural networks. The method introduces groundbreaking techniques for enhancing data privacy, demonstrated with comprehensive experimental results.

DAP is part of the AI-SPRINT project.

Prerequisites

• Python 3.x
• TensorFlow 2.x

Installation

To get started with DAP, follow these steps:

1. Clone the Repository:

git clone https://github.com/EugenioTL/discriminative-adversarial-privacy

2. Install Dependencies: Navigate to the project directory and install the required Python packages:

pip install -r requirements.txt

Usage

• Jupyter Notebook Demo: 'Discriminative Adversarial Privacy - Privacy Preserving Component Demo.ipynb' provides a comprehensive demonstration of DAP's capabilities. It's an excellent starting point for understanding the project's functionality.

• Running Scripts: The 'scripts' directory contains various scripts for tasks such as data preprocessing, model training, and evaluation. These scripts are vital for experimenting with and implementing the DAP framework in your projects.

• Model Directory: The 'models' directory houses the machine learning models used in DAP. These models are central to the framework's functionality and can be customized as per specific requirements.

Docker Deployment

For ease of deployment, especially in isolated environments, DAP can be containerized using Docker:

1. Build the Docker Image:

docker build -t dap-image .

2. Run the Container:

docker run -it dap-image

Citation

Should you find this repository useful, please consider citing:

@inproceedings{Lomurno_2023_BMVC,
author    = {Eugenio Lomurno and Alberto Archetti and Francesca Ausonio and Matteo Matteucci},
title     = {Discriminative Adversarial Privacy: Balancing Accuracy and Membership Privacy in Neural Networks},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year      = {2023},
url       = {https://papers.bmvc2023.org/0799.pdf}
}

License

This project is licensed under the terms of the LICENSE file. Please refer to the file for full licensing details.

Contact Information

For further inquiries or collaboration opportunities, feel free to contact the authors or raise an issue in this repository.