GAN-Based Tabular Synthetic Data

Description

GAN-based tabular synthetic data generation trains a generative model on a real table, then samples new rows from it to produce a synthetic table that can be shared or used in place of the original. In CTGAN (Conditional Tabular GAN), a generator network learns to produce rows and a discriminator network learns to tell them from real ones. CTGAN suits tables through mode-specific normalisation, which treats each numeric column as a mixture of clusters, and through training-by-sampling, which shows rare categories to the model often enough for it to learn them. TVAE is a related model built on a variational autoencoder, another kind of generative network, rather than a GAN. The synthetic table aims to preserve the original's column distributions and correlations, and its rows are generated rather than copied, but this is not a privacy guarantee, because a generator can memorise and reproduce training rows.

Example Use Cases

Privacy

Generating synthetic patient records from hospital databases using CTGAN to enable multi-institution collaborative research on treatment outcomes without sharing identifiable health data across organisational boundaries.

A bank's analytics team generates a synthetic copy of its customer transactions with CTGAN and gives it to an external vendor building a spending-insights dashboard. Before release, the team checks the table for rows that closely match real customers and decides whether it is safe to hand over.

A government statistics office produces a CTGAN synthetic version of a restricted household survey so researchers can develop and test analysis code before applying for access to the real data. The office uses it to decide what can be released openly and what stays in a secure facility.

Limitations

Training instability inherent to GAN architectures (mode collapse, vanishing gradients) can result in synthetic data that fails to capture the full diversity of the original distribution, particularly for rare categories or tail distributions.
There is no reliable rule for setting CTGAN's hyperparameters, such as epochs, batch size and learning rates, so users must experiment and watch whether the generator and discriminator losses converge. Training time grows with the number of epochs and can be shortened with a GPU.
Privacy guarantees are not inherent. Without additional mechanisms such as differential privacy, GAN-generated data may leak information about training records through memorisation or membership inference attacks.
CTGAN samples each row independently from fully connected networks, so it does not preserve row order, time dependencies or links between tables, and columns with many categories are harder to learn because each value is one-hot encoded.
The synthetic table can only reproduce what the training table contains: biases, gaps and thinly covered subgroups carry over or are lost, and models trained on synthetic rows are normally no better than models trained on the real table they imitate. Fidelity has to be tested separately against real data before the table is used.

Resources

Research Papers

Modeling Tabular data using Conditional GAN

Lei Xu et al.•Oct 28, 2019

Modeling the probability distribution of rows in tabular data and generating realistic synthetic data is a non-trivial task. Tabular data usually contains a mix of discrete and continuous columns. Continuous columns may have multiple modes whereas discrete columns are sometimes imbalanced making the modeling difficult. Existing statistical and deep neural network models fail to properly model this type of data. We design TGAN, which uses a conditional generative adversarial network to address these challenges. To aid in a fair and thorough comparison, we design a benchmark with 7 simulated and 8 real datasets and several Bayesian network baselines. TGAN outperforms Bayesian methods on most of the real datasets whereas other deep learning methods could not.

Software Packages

SDV

May 11, 2018

Synthetic data generation for tabular data

sdv-dev/CTGAN

Feb 25, 2026

Conditional GAN for generating synthetic tabular data.

Tutorials

How to Generate Real-World Synthetic Data with CTGAN | Towards ...

Miriam Santos•Apr 13, 2023

Documentations

CTGANSynthesizer | Synthetic Data Vault

Jun 12, 2026

Related Techniques

Name	Description	Assurance Goals
Fairness GAN	Fairness GAN is a pre-processing method that generates a synthetic training dataset in which decisions no longer track a protected attribute such as gender or skin tone. A generative adversarial network (GAN) is a pair of networks trained against each other, one producing fake examples and one trying to tell them from real ones. Here the generator is conditioned on the protected attribute and produces both an example and its decision label, such as hired or rejected (images in the original work, table rows in Xu et al.'s separate FairGAN). An extra network tries to predict the protected attribute from the generated output, and the generator is rewarded for defeating it, so favourable decisions become more similar across groups. The user can aim for demographic parity, meaning equal rates of favourable decisions, or for equality of opportunity, which applies the constraint only to favourable decisions.	Fairness
Sequential CART Synthesis	Sequential CART synthesis creates an artificial copy of a confidential microdata set (records about individual people or organisations) that can be released in place of the real data, in the form used by the synthpop package for R. Variables are synthesised one at a time in a chosen order: the first is generated by sampling with replacement from its observed values, and for each later variable a classification or regression tree (CART) is fitted on the real data to predict it from the variables earlier in the order. Each synthetic record is passed down that tree using its already-synthesised values, and the new value is copied from a randomly chosen real record in the leaf it reaches, so the joint distribution of the data is rebuilt as a chain of conditional distributions. Disclosure risk is limited by setting a minimum leaf size, smoothing synthesised continuous values and removing synthetic records that reproduce unique real ones, and several synthetic copies can be released so that analysts can allow for the extra uncertainty that synthesis adds.	Privacy
k-Anonymity	k-Anonymity is a property that a table of real records about people can be made to satisfy before it is released or shared, in which every record shares its values on the quasi-identifiers with at least k-1 other records, so anyone linking the table to another source on those columns always matches a group of at least k records, which is k people only when each person has one record. Quasi-identifiers are columns that are not names or ID numbers but could be found elsewhere and combined to single a person out, such as date of birth, sex and postcode, and because the data holder chooses them, the protection covers only those columns. To reach a chosen k, each quasi-identifier is coarsened along a hierarchy set in advance (exact age to five-year band, full postcode to district), and the values or records that would still sit in a group smaller than k are suppressed, with a search choosing the coarsening that loses least information. The output is the transformed table of real records made less specific, not artificial ones, together with the k it achieves and a measure of the detail lost. Unlike differential privacy, it gives no bound on what an observer can learn, because it blocks linking on the listed columns but does not stop someone learning a sensitive value that everyone in a group shares.	Privacy