CTGAN

To handle continuous columns with complex, non-Gaussian distributions (often multi-modal), CTGAN employs a variational Gaussian mixture model (VGM). This technique estimates the number of modes and fits a Gaussian mixture to each continuous column. Values are then represented as a one-hot vector indicating the sampled mode and a scalar value normalized within that mode. This prevents the discriminator from easily distinguishing between real and synthetic data based on simple statistical properties, ensuring high statistical fidelity.

Real tabular data often suffers from severe categorical imbalance, where certain discrete values appear infrequently. Standard GANs ignore these minority classes. CTGAN introduces training-by-sampling, which conditions the generator by sampling from a log-frequency distribution of categorical columns. This ensures that rare categories are represented proportionally during training, forcing the model to learn valid representations for tail-end data points rather than collapsing to majority modes.

The generator is conditioned using a conditional vector that specifies the desired discrete value for a selected column. This vector is concatenated with the latent noise input. To ensure the generator respects this condition, a cross-entropy loss is added to the standard adversarial loss, penalizing the generator if the synthetic row does not match the specified condition. This mechanism allows for controlled generation of specific data slices, directly addressing mode collapse in discrete spaces.

Unlike convolutional GANs designed for images, CTGAN uses a deep fully connected network with residual connections in both the generator and discriminator. This architecture is optimized for the mixed-type, non-spatial nature of tabular data. The residual blocks allow gradients to flow directly through the network, enabling the training of deeper models that can capture intricate, non-linear relationships between columns without suffering from vanishing gradients.

CTGAN is designed to natively synthesize datasets containing both continuous and categorical columns simultaneously. The discriminator receives a concatenated input of the mode-normalized continuous values and the one-hot encoded discrete values. This unified processing pipeline eliminates the need for separate models or manual feature engineering, allowing the adversarial game to learn the complex joint distribution between numerical ranges and discrete labels in a single end-to-end process.

To prevent the discriminator from easily memorizing the training data and to stabilize the adversarial game, CTGAN employs a PAC (Probably Approximately Correct) discriminator. Instead of evaluating single rows, the discriminator is fed a mini-batch of real or synthetic samples and must classify the entire batch as real or fake. This forces the discriminator to learn the joint distribution and inter-row variability, significantly reducing the risk of overfitting and improving the diversity of generated samples.

A critical failure mode in GAN training that CTGAN is explicitly designed to mitigate. Mode collapse occurs when the generator learns to produce only a limited variety of outputs that successfully fool the discriminator, ignoring entire modes of the real data distribution. In tabular data, this manifests as synthetic datasets that fail to represent minority classes or rare continuous value ranges. CTGAN combats this with conditional generation and training-by-sampling to enforce coverage of all categories.

The quantitative measure of how accurately a synthetic dataset preserves the properties of the original data. Key metrics include:

• Marginal distributions: Column-wise statistical similarity
• Joint distributions: Multi-column correlation preservation
• Column shapes: Matching non-Gaussian, multi-modal patterns High statistical fidelity is the primary objective of CTGAN, ensuring downstream machine learning models trained on synthetic data perform comparably to those trained on real data.

A mathematical framework that provides formal privacy guarantees by injecting calibrated noise into training processes. When combined with CTGAN, DP-SGD can be applied during training to bound the influence of any single record on the final model. This creates a differentially private CTGAN that generates synthetic data with provable privacy loss (ε), ensuring that the presence or absence of an individual in the training set cannot be reliably inferred from the synthetic output.

The gold-standard evaluation paradigm for measuring synthetic data utility. A machine learning model is trained exclusively on synthetic data generated by CTGAN and then evaluated on a held-out real test set. The performance gap between TSTR and a model trained on real data quantifies the utility loss. A small gap indicates high-quality synthesis that captures the predictive signal of the original data without memorizing individual records.

The statistical probability that an adversary can link synthetic records back to real individuals. Even though CTGAN learns a distribution rather than memorizing records, overfitting can cause the generator to reproduce near-exact copies of training samples. Evaluating distance to closest record (DCR) and membership inference attack susceptibility is essential to ensure the synthetic data provides meaningful privacy protection and does not inadvertently leak personally identifiable information.