A CVAE-GAN-based Approach to Process Imbalanced Datasets for Intrusion Detection in Marine Meteorological Sensor Networks

Abstract

In marine meteorological sensor networks (MMSN), there are massive data flows transmitted within numerous nodes, resulting in serious potential consequences once any anomalous traffic implied launches an attack. Therefore, accurate identification and fast response to abnormal traffic is vital for intrusion detection system (IDS) constructions. Dataset imbalances cause classification models to erroneously bias to normal traffic, significantly restricting IDS developments and applications. This paper proposes an approach to deal with dataset imbalances in intrusion detections. This approach mitigates dataset imbalance impacts on IDSs from the data perspective, which is liable to process the input data in classification models. In this approach, CVAE-GAN is adopted as the data generation module to synthesize specified minority class samples, thus reducing dataset imbalance rate. ordering points to identify the clustering structure (OPTICS) is taken as the denoising algorithm to remove outliers and decrease the overlap extent between majority classes. An experiment on NSL-KDD dataset demonstrates that the proposed method obtains a high-quality dataset with reasonable distribution. This approach improves the classifier's identification ability for potential anomalous traffic.