关键词: 异常检测, 生成对抗网络, Lindeberg-Feller中心极限定理, 假设检验

Abstract: In the fields of network maintenance and operation, it attracts much attention how to detect and prompt the network anomalies in time. Anomalous events are less in dataset than the normal ones, leading to the fact that it is difficult to use the two-class classifications for anomaly detection because of the imbalance of data labeled as normal or anomalous. Meanwhile, anomalous events are in various patterns and there is little prior information about the anomaly that the users are concerned with, therefore, it is necessary to model the normal data and use them for anomaly detection by comparing the received data with the normal model. Based on Lindeberg-Feller central limit theorem, a hypothesis test is designed to detect whether the data to be tested is anomalous or not, according to the refusing area calculated by the confidential parameter. Finally, the theorem of this algorithm is simulated and the performance is also tested both on the common and the actual datasets. When the users take the correlation features of the anomalous events as the algorithm input, the recall ratio reaches 90%.

Key words: anomaly detection, generative adversarial networks, Lindeberg-Feller central limit theorem, hypothesis test