TW-FCM: An Improved Fuzzy-C-Means Algorithm for SPIT Detection

With the popularity of VoIP (Voice over Internet Protocol) systems, there has been an explosive growth in VoIP spam. In order to effectively prevent spam calls, various methods based on the analysis of call behavior features have been proposed. However, few of the existing methods consider that different features have different weights, resulting in a low detection precision of SPIT (Spam over Internet Telephony) users. Meanwhile, most methods are tested based on the experimental data generated by simulation, it is not sure whether these methods work well in the real world. In this paper, we propose a Weighted-Fuzzy-C-Means (W- FCM) algorithm, which can automatically adjust the weight of each call feature in the clustering process. Experiments based on the real world data show that our proposed algorithm could effectively improve the detection precision (about 6.7%) and recall (about 0.3%) of SPIT users. We also analyze the impact of different membership thresholds on the clustering results and propose a Threshold-W-FCM (TW-FCM) algorithm, through which we can select appropriate membership thresholds to alleviate the class-imbalance problem, and thereby improve the overall performance of SPIT detection compared with traditional FCM method.