CasTGAN: Cascaded Generative Adversarial Network for Realistic Tabular Data Synthesis

TL;DR

CasTGAN tackles the challenge of realistic tabular data synthesis with strong validity constraints and privacy considerations. It introduces a cascaded GAN where each feature is generated by its own generator and guided by auxiliary learners trained on the rest of the data, enabling better capture of feature interdependencies. The approach leverages Variational Gaussian Mixture modeling for numerical features, one-hot encoding for categoricals, and a WGAN-GP training objective, achieving strong machine learning utility while reducing invalid records compared to prior tabular GANs. The authors also explore white-box privacy attacks and show that perturbing auxiliary learners can boost robustness with limited sacrifice to data realism. Collectively, CasTGAN advances practical synthetic tabular data generation for cross-organization knowledge sharing under privacy constraints.

Abstract

Generative adversarial networks (GANs) have drawn considerable attention in recent years for their proven capability in generating synthetic data which can be utilised for multiple purposes. While GANs have demonstrated tremendous successes in producing synthetic data samples that replicate the dynamics of the original datasets, the validity of the synthetic data and the underlying privacy concerns represent major challenges which are not sufficiently addressed. In this work, we design a cascaded tabular GAN framework (CasTGAN) for generating realistic tabular data with a specific focus on the validity of the output. In this context, validity refers to the the dependency between features that can be found in the real data, but is typically misrepresented by traditional generative models. Our key idea entails that employing a cascaded architecture in which a dedicated generator samples each feature, the synthetic output becomes more representative of the real data. Our experimental results demonstrate that our model is capable of generating synthetic tabular data that can be used for fitting machine learning models. In addition, our model captures well the constraints and the correlations between the features of the real data, especially the high dimensional datasets. Furthermore, we evaluate the risk of white-box privacy attacks on our model and subsequently show that applying some perturbations to the auxiliary learners in CasTGAN increases the overall robustness of our model against targeted attacks.

Figures (7)

• Figure 1: The model architecture of CasTGAN. The cascade $\vec{G}$ is composed of generators $G_{1}, G_{2}, \ldots, G_{M}$ sequentially lined up. The auxiliary learners $AL_{1}, AL_{2}, \ldots, AL_{M}$ are fitted on the real data, and are utilised by their respective generators for querying the generation of the data features $m_{1}, m_{2}, \ldots, m_{M}$. The cascade of generators $\vec{G}$ takes noise vector $z$ as input, while the discriminator $D$ is trained to distinguish the real data from the synthetic data. As depicted, generators $G_{1}, G_{2}, \ldots, G_{M-1}$ receive three losses: loss directly from the discriminator, loss backpropagated from the previous generator and the loss from the auxiliary learner. Meanwhile, $G_{M}$ is passed the loss from its auxiliary learner and the loss from the discriminator.
• Figure 2: A close in visualization of generator $G_{i}$ in our GAN architecture. The depiction in the figure is applicable to all generators in the cascaded layout, except for $G_{1}$, which receives only noise vector $z$ as input.
• Figure 3: A detailed depiction of auxiliary learner $AL_{i}$ in CasTGAN. The training of the LightGBM model on the real data occurs prior to the training of the GAN model. Meanwhile, the synthetic data from the generator $G_{i}$ is queried against the auxiliary learner $AL_{i}$ during the training iterations of the GAN to compute the auxiliary loss of the generator's target feature.
• Figure 4: Discrete and continuous univariate features distribution plots for the Bank dataset.
• Figure 5: Adult dataset correlation map plots for the real and the synthetic data generated by CasTGAN. Larger absolute values indicate a stronger correlation, either positively or negatively.