(ICWS 2022) Towards An Improved Bi-GAN-Based End-to-End One-Class Classifier For Anomaly Detection In Cloud Data-Centers

Jiale Zhao (ChongQing University, China); Peng Chen and Juan Chen (Xihua University, China); Xianhua Niu (Xihua University); Yunni Xia (Chongqing University, China)

With the continuous popularization of cloud computing services, ensuring the stable operation of cloud data-centers has become a hot research issue. Efficient anomaly detection for a running cloud data-center and timely judgment of the cause of anomalies will help to fundamentally improve the reliability of cloud data-center infrastructures. Nevertheless, traditional anomaly identification approaches are challenging to meet the requirements of cloud data-centers with increasingly complex system structures. Due to the low feasibility of labeling data, machine learning methods based on supervised learning also make it difficult to perform efficient anomaly detection in cloud data-centers. This work exploits novel GAN-based generative models and end-to-end one-class classification for optimizing unsupervised anomaly identification. A new Bi-GAN-based Heterogeneous Anomaly-reconstruction One-class Classifier (BG-HA-OC) is developed optimizing a one-class classifier and an anomaly scoring function. The Generator-Encoder-Discriminator Bi-GAN is capable of performing practical anomaly score computation and capturing fine temporal features. In the empirical study, we demonstrate that our proposed framework outperforms its peers upon third-party anomaly detection methods on anomaly benchmarks and synthetic datasets.