使用圖神經網路偵測 PTT 的低活躍異常帳號;Using Graph Neural Networks to Detect Inactive Spammers on PTT

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/89769

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/89769

Title:	使用圖神經網路偵測 PTT 的低活躍異常帳號;Using Graph Neural Networks to Detect Inactive Spammers on PTT
Authors:	王瑞緣;WANG, JUI-YUAN
Contributors:	資訊工程學系
Keywords:	網軍;異常帳號;低活躍異常帳號;低活躍帳號;圖神經網路;批踢踢;批踢踢實業坊;spammer;social networks;inactive user;inactive spammer;graph convolutional networks;PTT
Date:	2022-07-19
Issue Date:	2022-10-04 11:58:50 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著社群媒體的興起，雇用「公關公司」在網路上散播不實消息，成為左右時事輿論的新興手法，公關公司的大量帳號，常被各大論壇視為異常帳號。國內外皆有學者以深度學習偵測異常帳號，但我們發現，現階段偵測異常帳號的論文中，並沒有針對帳號的活躍程度作探討。本篇論文中，我們依據帳號在限定時間內的活動次數定義出「活躍值」的概念，我們觀察到，用簡單的卷積類神經網路 (Convolutional Neural Network) 模型，即可在偵測高活躍異常帳號的任務中達到 0.9169 的 ROC 曲線下的面積 (AUROC)，但是偵測低活躍的異常帳號卻只有 0.7830，顯示出偵測低活躍異常帳號是非常棘手的任務。我們利用使用者與使用者之間的關係建立社群網路，以提供額外的特徵作為訓練的資料，並引入圖神經網路，成功改善偵測低活躍異常帳號的任務。;With the rise of social media, hiring public relations companies to spread fake news on the Internet has become an emerging method to manipulate public opinions. These large number of accounts owned by public relations companies are regarded as spammers by most online forums. Researchers have used deep learning techniques to detect abnormal accounts. However, we found that these studies likely conducted experiments mainly on the active users. In this thesis, we define the concept of ”Active Value” based on the number of activities of an account within a unit period. For active users, even a simple Convolutional Neural Network model can distinguish a spammer from a regular user: the area under the ROC curve (AUROC) achieves 0.9169. However, for the inactive users, the score drops to 0.7830. The result indicates that detecting inactivity spammers is much more challenging. We use user-to-user relationships to build a social network. We apply graph neural networks to the social network and extract additional social features as training clues. Experimental results show that these strategies better distinguish the spammers from regular users, especially when these users have limited activities.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	55	View/Open

社群 sharing

Loading...