On Identifying Suspicious Accounts Using Anomaly Detection Technology

Abstract

In recent years, threats such as "fake news" and "disinformation" have reached the level of national security in information warfare, and have become an important research issue. For example, as early as 2014, Russia intervened to influence Ukraine’s Crimea referendum, and in the recent Ukrainian-Russian War, we can see that in many communities, whether Russia or the others, the media takes the wind. This article focuses on the accounts and posts that publish suspicious information, and uses Twitter’s official project website — Transparency website. Twitter defines suspicious accounts as accounts that manipulate disinformation related to the government or state, and publishes them after investigation and confirmation. Different from the previous identification methods, in this paper we use the "anomaly detection" technology in machine learning to train a classifier that can distinguish abnormal messages and abnormal accounts with high accuracy. For the dataset, we established a data crawling system based on the ETL framework, and crawled official accounts and tweets of celebrities. And use the normal posts posted by the accounts with blue tick, whose identities have been officially confirmed, to verify the performance of the classifier. From the experimental results, we found that the accuracy of our identification method reached 96%.