Applying Deep Learning to Balancing Network Intrusion Detection Datasets
Po‐Jen Chuang, Dong-Ye Wu · 2019
In this investigation, we apply deep learning to generate a desirable data generation model which helps to balance the network intrusion detection datasets and enhance the detection performance. The basic idea is to adopt deep variational autoencoders to generate new data and adjust the imbalance of datasets into more favorable balanced datasets. We then use the generated balanced datasets to reduce the deviation of classifiers (incurred by data imbalance) in training and, as a result, to enhance the performance of intrusion detection. As experimental results exhibit, in comparison to unbalanced datasets, balanced datasets can practically lead to better classification accuracy, especially when facing unknown attacks. Balanced datasets also help to solve the over-fitting problem (due to imbalance of datasets) in intrusion detection model training, to ensure that the trained intrusion detection model will not misjudge new types of data - even if they are not in the training dataset.