[Paper Review] Fed-TGAN: Federated Learning Framework for Synthesizing Tabular Data

Zilong Zhao, Robert Birke+2 more|arXiv (Cornell University)|Aug 18, 2021

Privacy-Preserving Technologies in Data25 references25 citations

TL;DR

Fed-TGAN proposes a federated learning framework for tabular GANs with privacy-preserving encodings and table-similarity weighted aggregation, achieving faster convergence and higher data similarity than baselines.

ABSTRACT

Generative Adversarial Networks (GANs) are typically trained to synthesize data, from images and more recently tabular data, under the assumption of directly accessible training data. Recently, federated learning (FL) is an emerging paradigm that features decentralized learning on client's local data with a privacy-preserving capability. And, while learning GANs to synthesize images on FL systems has just been demonstrated, it is unknown if GANs for tabular data can be learned from decentralized data sources. Moreover, it remains unclear which distributed architecture suits them best. Different from image GANs, state-of-the-art tabular GANs require prior knowledge on the data distribution of each (discrete and continuous) column to agree on a common encoding -- risking privacy guarantees. In this paper, we propose Fed-TGAN, the first Federated learning framework for Tabular GANs. To effectively learn a complex tabular GAN on non-identical participants, Fed-TGAN designs two novel features: (i) a privacy-preserving multi-source feature encoding for model initialization; and (ii) table similarity aware weighting strategies to aggregate local models for countering data skew. We extensively evaluate the proposed Fed-TGAN against variants of decentralized learning architectures on four widely used datasets. Results show that Fed-TGAN accelerates training time per epoch up to 200% compared to the alternative architectures, for both IID and Non-IID data. Overall, Fed-TGAN not only stabilizes the training loss, but also achieves better similarity between generated and original data. Our code is released at https://github.com/zhao-zilong/Fed-TGAN.

Motivation & Objective

• Enable training CTGAN-style tabular GANs across decentralized clients without sharing raw data.
• Develop privacy-preserving column encodings to initialize a global encoder without exposing private data.
• Design a table-similarity aware weighting scheme to aggregate local models under non-IID tabular data.
• Demonstrate faster convergence and higher similarity between generated and real data compared with baselines across multiple datasets.

Proposed method

• Adopt a federated learning structure with a federator and multiple clients where each client trains a local generator and discriminator.
• Implement privacy-preserving feature encoding by collecting column statistics (categorical frequencies and continuous VGM parameters) to build global encoders without accessing raw data.
• Distribute global encoders LE_j and VGM_j to clients to initialize models with unified input/output structures.
• Compute per-client aggregation weights using per-column divergences (Jensen-Shannon for categorical, Wasserstein for continuous) and data quantity, then apply a softmax-normalized weighted aggregation.
• Implement the system using PyTorch RPC for asynchronous training rounds and centralized aggregation, addressing GPU-CPU data transfer limitations.

Experimental results

Research questions

• RQ1 Can Fed-TGAN learn tabular GANs from decentralized data while preserving privacy?
• RQ2 Does privacy-preserving initialization and table-aware weighting improve convergence speed and data realism under IID and non-IID conditions?
• RQ3 How do Fed-TGAN’s results compare to centralized CTGAN, vanilla FL, and multi-discriminator MD-GAN across standard tabular datasets?

Key findings

• Fed-TGAN reduces training time per epoch by up to 200% compared to the multi-discriminator baseline.
• Fed-TGAN achieves higher similarity (lower Avg-JSD and Avg-WD) to real data than baselines across four datasets in the ideal IID case.
• Under imbalanced data across clients, Fed-TGAN converges faster than vanilla FL-TGAN.
• Ablation shows that removing table-similarity weighting degrades performance, confirming its usefulness.
• Fed-TGAN maintains stable convergence and improved similarity under non-IID distributed data.