HT-Fed-GAN: Federated Generative Model for Decentralized Tabular Data Synthesis.

Area of Science:

• Computer Science
• Data Privacy
• Machine Learning

Background:

• Existing privacy-preserving data synthesis (PPDS) methods struggle with decentralized tabular data due to mixed data types and imbalanced distributions.
• Current federated generative models are primarily designed for image or text data, not the complexities of tabular datasets.

Purpose of the Study:

• To propose a novel federated generative model, HT-Fed-GAN, for privacy-preserving tabular data synthesis in a distributed multi-party environment.
• To address the limitations of existing methods in handling multimodal distributions and imbalanced attributes in decentralized tabular data.

Main Methods:

• Developed HT-Fed-GAN, incorporating a federated variational Bayesian Gaussian mixture model (Fed-VB-GMM) for multimodal distributions.
• Implemented federated conditional one-hot encoding with conditional sampling for categorical attribute representation and rebalancing.
• Utilized a privacy consumption-based federated conditional GAN for privacy-preserving decentralized data modeling.

Main Results:

• HT-Fed-GAN demonstrated the optimal balance between data utility and privacy levels across five real-world datasets.
• Generated tables using HT-Fed-GAN exhibited superior statistical similarity to original tables.
• Evaluation scores confirmed HT-Fed-GAN's outperformance against state-of-the-art models in machine learning tasks.

Conclusions:

• HT-Fed-GAN effectively addresses the challenges of privacy-preserving tabular data synthesis in decentralized settings.
• The proposed model offers significant improvements in both data utility and privacy preservation compared to existing approaches.
• HT-Fed-GAN represents a significant advancement for secure and effective analysis of distributed tabular data.