Data imbalance in network intrusion detection datasets tends to incur underfitting or deviation in classifier training. This investigation applies Batched Variational AutoEncoders (B-VAE) to generate a desirable data generation model which can balance intrusion detection datasets to enhance the detection practice. To improve insufficient VAE decoder training in the VAE approach, we apply B-VAE to train one decoder for each piece of data by a batched duplicated data and form multiple batched VAEs to provide sufficient decoder training. The unique practice of B-VAE makes the generated data all similar to but different from the original data, to secure desirable data balance for better classifier training and classification results. Experimental evaluation conducted to compare the performance of related balancing approaches shows that our B-VAE outperforms others in that it is able to maintain the same classification accuracy (in terms of F1-scores) regardless of any Imbalance Ratio (IR) change. Specifically, B-VAE manages to solve the problem of insufficient decoder training in existing approaches and so to enhance the intrusion detection performance—mainly because it can secure balanced data generation to lift the classification accuracy due to sufficient decoder training and utilization of exact features.