Detection of Change of Users in SNS by Two Dimensional CNN

Abstract

In this paper, we proposed a method for detecting hacked accounts in SNS without predetermined features since trend of topics and slang expressions always change and hackers can make messages which are matched with the predetermined features. There are some researches in which a hacked account or impersonation in SNS is detected. However, they have problems that predetermined features were used in their method or evaluation procedure was not appropriate. On the other hand, in our method, a feature named 'category' is automatically extracted among recent tweets by machine learning. We evaluated the categories with 1,000 test accounts. As a result, 74.4% of the test accounts can be detected with the rate up to 96.0% when they are hacked and only one new message is posted. Moreover, 73.4% of the test accounts can be detected with the rate up to 99.2% by one new posted message. Furthermore, other hacked accounts can also be detected with the same rate when several messages are sequentially posted.