Anomaly detection is a process to identify abnormal behavior that does not confirm the normal behavior. The abnormal behavior clues are few because it appears rarely. To detect each abnormal behavior, the problem is transformed into a multi-class classification task where it lies into data imbalance representation. In data imbalance setting, minority classes are over-sampled to improve the performance of the classifier. Existing methods are unable to learn the distribution of the minority class and effects the performance of classifier. In this research, we introduced a data generative model (DGM) to improve the minority class presence in the anomaly detection domain. Our approach is based on a conditional generative adversarial network to generate synthetic samples for minority classes. It includes the KL-divergence to guide the model towards the true learning of minority class distribution. In this way, model learns the complex underlying data distribution and generates new samples. We performed experiments over the two benchmark datasets NSL-KDD and UNSW-NB15 that are publicly available to demonstrate the effectiveness of our approach. Furthermore, the comparative analysis with the existing approaches confirms the stability and superiority of our presented model.