Predicting Malicious Users on Anonymous Chat Networks

Malicious users on chat network systems would reduce the willingness of benevolent users to chat on the same network. Thus it is often desirable to classify malicious users based on their personal profile and chat contents. In this study, different algorithms including Naive Bayes, SVM, Decision Table, Multilayer Perceptron, and Logistic classification are applied on the dataset from an anonymous chat network Chatous. It is found that by using both empirical features and chat word features together with the SVM algorithm, the best Fscore achieved is 84.9%. This result shows the possibility of classifying malicious users from benevolent users on a chat network and may lead to chat quality improvement such as by restricting malicious user from chatting with benevolent users. Keywords—Malicious users detection, Word feature extraction, Support Vector Machine, Machine learning, Chat Network.