SPIT callers detection with unsupervised Random Forests classifier

As VoIP (Voice over IP) grows rapidly, it is expected to prevail tremendous unsolicited advertisement calls, which type of calls is referred to SPIT (SPam over Internet Telephony). SPIT detection is more difficult to execute than email SPAM detection since the callee or SPIT detection system does not tell whether it is SPIT or legitimate call until he/she actually takes a call. Recently, many SPIT detection techniques are proposed by finding outliers of call patterns. However, most of these techniques suffer from setting a threshold to distinguish that the caller is legitimate or not and this could cause to high false negative rate or low true positive rate. This is because these techniques analyse call pattern by a single feature e.g. call frequency or average call duration. In this paper, we propose a multi-feature call pattern analysis with unsupervised Random Forests classifier, which is one of the excellent classification algorithms. We also propose two simple but helpful features for better classification. We show the effectiveness of Random Forests based classification without supervised training data and which features contribute to classification.