MetaCom: Profiling Meta Data to Detect Compromised Accounts in Online Social Networks

Abstract

Social networks have become the center of research and its increasing popularity has also led to its misuse by a number of malicious users. In order to conduct various malicious activities on the online platforms, malevolent users rely on spam, fake or compromised accounts to disseminate their illegitimate information. This paper addresses the detection of compromised accounts so that the concerned user can take the necessary action to mitigate the effect of compromise. Unlike most of the existing techniques where text based features are used to address the problem, this research examines the efficiency of meta data information associated with each text in detecting the compromised accounts. Secondly, we have studied the problem from both unary as well as binary classification perspectives where efficiency of respective machine learning classifiers have been analyzed on the basis of different evaluation metrics. Amongst five binary classifiers, Random Forest attained highest efficiency achieving 92.66% F-score. On the other hand, with one class classifiers, OCC-SVM with rbf kernel attained maximum performance with an average F-score of 72.72%.