Power-Learning: Differentially Private and Model Agnostic Tabular Data Embeddings

This repository contains the official implementation of "Power Learning: Differentially private and model agnostic tabular data embeddings", which introduces a novel framework for sharing neural network activations with formal differential privacy guarantees.

Overview

Traditional collaborative learning approaches focus on sharing model weights between clients and servers. Power-Learning introduces a new paradigm that enables:

• Privacy-preserving embedding generation from tabular data
• Model-agnostic server processing (compatible with neural networks, random forests, XGBoost)
• Formal differential privacy guarantees
• Efficient single-round communication
• Reduced client-side computation

Repository Structure

├── Adult_income/        # Adult Income dataset experiments
├── Churn/              # Churn dataset experiments  
├── Forrest_Cover_Type/ # Forest Cover Type dataset experiments
├── Higgs/             # Higgs dataset experiments
├── LLM/               # Language model components
├── Rossman/           # Rossman dataset experiments
├── _pycache_/         # Python cache files
└── .gitignore         # Git ignore file

Requirements

• Python 3.x
• PyTorch
• NVIDIA GPU with CUDA support
• Additional requirements listed in requirements.txt (to be added)

Key Components

1. Client-Side Processing:

   • Privacy-inducing training
   • Privacy level calibration
   • Lightweight model architecture

2. Server-Side Processing:

   • Flexible model choice (Neural Networks, Random Forests, XGBoost)
   • High computational capacity utilization
   • Model-agnostic processing

Datasets

The implementation has been tested on three main datasets:

1. Forest Cover Type

   • 580k samples
   • 54 features
   • 7 classes

2. Higgs Boson

   • 240k samples
   • 30 features
   • 2 classes

3. Adult Income

   • 48k samples
   • 14 features
   • 2 classes

Usage

(To be added: Instructions for running the code, including example commands and parameter settings)

Results

Our method demonstrates:

• Better privacy-utility trade-offs compared to DP-Adam
• Significant computational efficiency improvements
• Robust defense against feature space hijacking attacks
• Only 0.36% information leakage compared to 4.5% in baseline methods

Citation

If you find this work useful in your research, please consider citing:

(Citation to be added after publication)

License

This project is licensed under [LICENSE_NAME] - see the LICENSE file for details.

Hardware & Code

The experiments were conducted using:

• NVIDIA A100-SXM4-80GB GPU
• Python implementation with PyTorch

Contributing

We welcome contributions to this repository. Please feel free to submit pull requests or create issues for bugs and feature requests.

Contact

For questions and feedback, please create an issue in the repository.

Acknowledgments

(To be added: Any acknowledgments to contributors, funding sources, etc.)