Cyber Threat Detection Based On Angle-Based Subspace Anomaly For Machine Learning Applications

Abstract

In real time applications related to social media, conventional anomaly detection techniques are not applicable as the accuracy is degraded due to higher dimensionality thereby hampering these applications leading to cyber threat. This paper focuses on an approach that can select feature subspaces of social media which have meaningful information and thereby conduct anomaly detection in the projected subspace correspondingly. Major goal is to maintain the accuracy of detection in the circumstance of high dimensionality detecting cyber threat. This approach determines the angle between any of the two lines for one of the anomaly candidate specifically where the first line is in connection with relevant data points along with centers of adjacent points and the other line is any parallel axis line. For a particular candidate in social media, any dimension having smallest angle with the first line is chosen as subspace parallel to the axis. In the projected subspace, the local outlierness is measured for an object by introducing normalized values of Mahalanobis distance. Artificial datasets are constructed for comparing proposed approach of detecting cyber threats in social media comprehensively and found to be accurate for machine learning applications.