Research on unsupervised anomaly data detection method based on improved automatic encoder and Gaussian mixture model
Xiangyu Liu, Shibing Zhu, Fan Yang, Shengjun Liang
Abstract
Abstract With the development of cloud computing, more and more security problems like “fuzzy boundary” are exposed. To solve such problems, unsupervised anomaly detection is increasingly used in cloud security, where density estimation is commonly used in anomaly detection clustering tasks. However, in practical use, the excessive amount of data and high dimensionality of data features can lead to difficulties in data calibration, data redundancy, and reduced effectiveness of density estimation algorithms. Although auto-encoders have made fruitful progress in data dimensionality reduction, using auto-encoders alone may still cause the model to be too generalized and unable to detect specific anomalies. In this paper, a new unsupervised anomaly detection method, MemAe-gmm-ma, is proposed. MemAe-gmm-ma generates a low-dimensional representation and reconstruction error for each input sample by a deep auto-encoder. It adds a memory module inside the auto-encoder to better learn the inner meaning of the training samples, and finally puts the low-dimensional information of the samples into a Gaussian mixture model (GMM) for density estimation. MemAe-gmm-ma demonstrates better performance on the public benchmark dataset, with a 4.47% improvement over the MemAe model standard F1 score on the NSL-KDD dataset, and a 9.77% improvement over the CAE-GMM model standard F1 score on the CIC-IDS-2017 dataset.