Learning from limited data remains a fundamental challenge in machine learning. While generative adversarial networks offer a promising solution by synthesizing training instances, they often produce biased distributions when trained on small datasets. We propose EnhanceGAN, a novel framework that addresses this limitation through two key innovations: an ensemble-based discriminator correction mechanism and a Markov chain Monte Carlo sampling method for generator refinement. Furthermore, we introduce a two phase training strategy where downstream classifiers are first pre-trained on generated data, then fine-tuned on real instances to mitigate synthetic data bias. Comprehensive experiments across multiple datasets demonstrate that EnhanceGAN is effective in data-scarce scenarios.