ppml 0.1.2

A privacy-preserving machine learning package

🥯 Privacy-Preserving Machine Learning (PPML)

This repository implements privacy-preserving machine learning techniques, starting with a differentially private GAN for image generation. Our goal is to develop and showcase various PPML methods that enable data analysis and model training while protecting individual privacy.

Our blog post explains the importance of privacy in machine learning and introduces our first implementation. 🥯

🥯 Features

• Privacy-preserving machine learning techniques
• Differential privacy implementations
• PyTorch and Opacus integration

🥯 Current Implementations

1. Differentially Private GAN (DP-GAN) for image synthesis
   • Based on Radford et al. (2015) and Xie et al. (2018)
   • Convolutional GAN architecture
   • Customizable privacy budget (ε)

🥯 Applications

Our PPML techniques can be applied to various scenarios:

• Secure data sharing for research
• Privacy-preserving model training
• Confidential data analysis in sensitive domains (e.g., healthcare, finance)
• Building privacy-respecting AI systems

🥯 Setup

1. Clone the repository:

   git clone https://github.com/your-username/PPML.git
   cd PPML
   

2. Install dependencies:

   pip install -e .
   

🥯 Usage

DP-GAN for Image Synthesis

1. Place your dataset in the data folder. The default configuration expects a folder structure similar to the CelebA dataset.

2. Open dp.py and set the dataroot variable to point to your dataset folder:

   dataroot = "path/to/your/dataset"
   

3. (Optional) Adjust hyperparameters in dp.py:

   • EPSILON: Privacy budget (default: 50.0)
   • batch_size: Number of images per batch (default: 128)
   • num_epochs: Number of training epochs (default: 5)
   • lr: Learning rate (default: 0.0002)

4. Run the training script:

   python dp.py
   

5. Monitor the training progress. The script will print loss values and privacy guarantees every few iterations.

6. After training, find the trained Generator model saved as netG_dpgan.pth in the project directory.

7. To generate new images using the trained model, create a new script that loads the saved model and feeds random noise through it.

🥯 Example

We trained the DP-GAN on CelebA with ε=50. Results:

🥯 Roadmap

We plan to expand this repository with more PPML techniques, including:

• Federated Learning implementations
• Secure Multi-Party Computation (SMPC) for distributed machine learning
• Homomorphic Encryption-based machine learning models
• Privacy-preserving data analysis tools

Stay tuned for updates!

🥯 Contributing

We welcome contributions to improve and expand this project! Here are some guidelines to get started:

1. Fork the repository and create your branch from main.
2. If you've added code that should be tested, add tests.
3. Ensure your code passes all tests.
4. Make sure your code lints.
5. Issue a pull request with a comprehensive description of changes.

For more detailed information, please read our Contribution Guidelines.

🥯 License

MIT License

🥯 Acknowledgements

• Radford et al. (2015): DCGAN architecture
• Xie et al. (2018): Differentially private GAN
• Opacus: Differential privacy library

🥯 Contact

Open an issue or email [[e-mail]].