Empirical Evaluation and New Design for Fighting Evolving Twitter Spammers

To date, as one of the most popular online social networks (OSNs), Twitter is paying its dues as more and more spammers set their sights on this microblogging site. Twitter spammers can achieve their malicious goals such as sending spam, spreading malware, hosting botnet command and control (C&C) channels, and launching other underground illicit activities. Due to the significance and indispensability of detecting and suspending those spam accounts, many researchers along with the engineers at Twitter Inc. have devoted themselves to keeping Twitter as spam-free online communities. Most of the existing studies utilize machine learning techniques to detect Twitter spammers. “While the priest climbs a post, the devil climbs ten.” Twitter spammers are evolving to evade existing detection features. In this paper, we first make a comprehensive and empirical analysis of the evasion tactics utilized by Twitter spammers. We further design several new detection features to detect more Twitter spammers. In addition, to deeply understand the effectiveness and difficulties of using machine learning features to detect spammers, we analyze the robustness of 24 detection features that are commonly utilized in the literature as well as our proposed ones. Through our experiments, we show that our new designed features are much more effective to be used to detect (even evasive) Twitter spammers. According to our evaluation, while keeping an even lower false positive rate, the detection rate using our new feature set is also significantly higher than that of existing work. To the best of our knowledge, this work is the first empirical study and evaluation of the effect of evasion tactics utilized by Twitter spammers and is a valuable supplement to this line of research.