Iterative Anomaly Detection Algorithm Based on Time Series Analysis

Abstract

Anomaly detection is a concerned field in recent years, which plays an important role in improving network performance, finding network anomalies in time and ensuring network security. Traditional anomaly detection algorithms are based on whether the traffic features exceed a given threshold to determine the anomaly, which have low accuracy. And recently, lots of works which are based on statistical model or machine learning can't deal with the impact of outliers on the subsequent fitting, so the results are not perfect enough. In this paper, a new iterative anomaly detection algorithm based on time series analysis is proposed. The algorithm detects anomalies by automatically fitting the best ARMA model iteratively, and detects the first anomaly point in each iteration. This method can produce more precise results, and has the features of high accuracy and low misjudgment rate.