有鑒於社群媒體傳播的重要性與影響與日俱增,逐漸形塑出我們對於整個世界的認知,在社群媒體上也盛行利用大量網軍及假帳號的協同性造假行為來帶動輿論風向,而其中針對社群媒體帳號遭盜用或是帳號遭盜用後被拍賣牟利這類情形,就屬於盜用帳號偵測的問題。 本論文旨在研究社群媒體上盜用帳號的探勘方法。本論文所提出的方法主要運用盜用可疑時間點偵測。第一階段先根據帳號的活動量、活動時間等時間行為特徵,結合統計學領域的變化點偵測,進行可疑時間點偵測,以找到帳號行為改變的可疑時間點。第二階段,針對可疑時間點數量,以及每個可疑時間點前後的時間行為特徵、空間行為特徵、文字內容、立場光譜等9項特徵是否有明顯差異變化進行檢驗。最後,則透過模擬生成盜用帳號方法,去驗證預測模型之效果。 最後,我們先以Twitter為例,證明在真實資料集狀況下,我們的研究方法也能有較好的表現。接著以PTT人工生成資料方式,驗證運用可疑時間點偵測及所提出的行為特徵之效果。
In view of the increasing importance and influence of social media communication, which has gradually shaped our cognition of the whole world, it is also become popular in social media to use a large number of cyber warriors and fake accounts to manipulate public opinion. Social media accounts that are hijacked or auctioned for profit is considered as the compromised accounts. This study aims to research the detection method of social media compromised account. In the first step, based on the activity volume and activity time of the account, we propose the integration of change point detection algorithm to find the suspicious time point when the behavior of account changes. In the second step, we propose the interest feature, the polarity feature and the change point feature and examine whether there is major change in the account's behavioral feature before and after the suspicious time point. Finally, the prediction model is validated by simulating the generation of compromised accounts. We verify the proposed change point approach by using Twitter Dataset. Then, we verify the performance of the proposed features by using synthesized PTT dataset. The experiments show that the proposed approach performs better than existing research.