Defence organisations collect operational, medical, cyber, logistics, and maintenance records that are valuable for model development but difficult to share because they may contain sensitive or mission-revealing information. Synthetic tabular data offers a practical way to support experimentation, benchmarking, and training while reducing direct exposure of original records. This paper presents an original review of Generative Adversarial Network (GAN)-based approaches for tabular data synthesis, with emphasis on the requirements of defenceoriented workflows. It explains the adversarial objective, the Wasserstein formulation with gradient penalty, and preprocessing methods that allow neural generators to handle numerical and categorical fields. The paper also discusses evaluation through fidelity, downstream utility, robustness, fairness, and privacy testing. Rather than treating synthetic data as automatically safe, the analysis argues for a documented validation pipeline that measures both model performance and disclosure risk before synthetic records are released or used in operational decision support.