假的真不了！基於機器學習綜合個人特徵、貼文特徵與圖像特徵之臉書非真實帳號辨識模型

Undeniably Inauthentic ! Facebook Inauthentic Account Detection Model Based on Machine Learning Integrating Personal Characteristics, Post Features, and Image Attributes

莊玉馨(YU-HSIN CHUANG)

指導教授 ： 金志聿

中原大學/商學院/資訊管理學系/碩士(2024年)

https://doi.org/10.6840/cycu202400106

摘要

在當代資訊傳播領域中，社群媒體扮演著至關重要的角色。人們高度仰賴社群媒體獲取資訊，促使社群媒體遂成個人、企業或品牌進行資訊曝光與建立口碑的關鍵渠道。然而，在高度依賴社群媒體競爭的背景下，「非真實帳號」問題逐漸浮出檯面，更對社會造成了嚴重的危害；此外，在台灣的社群媒體使用者中，有超過七成的用戶經常使用臉書；因此，針對臉書社群中的非真實帳號進行有效辨識顯得尤為重要。
近年來，偵測社群媒體中非真實帳號的辨識研究已從早期使用單一數據特徵轉而結合多種特徵和不同模態之方法，透過增加特徵維度提高非真實帳號的辨識準確性。此外，過去非真實帳號辨識研究主要聚焦Twitter平台，在臉書平台的相關研究較少，國內的相關研究更是寥寥無幾；本研究提出較過往文獻更多維的特徵屬性，包含臉書用戶個人帳號資料、貼文資料，並透過特徵工程方法萃取包含用戶的個人資料特徵、貼文行為特徵以及透過圖形模態分析的圖解特徵等逾四十種特徵，應用隨機森林、支援向量機以及堆疊式集成學習方法進行建模預測，辨識臉書社群之非真實帳號。
本研究實驗發現：一、隨機森林與集成學習演算法之非真實帳號辨識模型辨識準確率達99%，展現實務應用價值。二、本研究發現結合社群互動特徵進行非真實帳號辨識，能夠大幅提升模型準確率。三、透過探索性分析發現圖像辨識與實際資訊之差異、貼文留言平均數和貼文打卡與標記的占比分析，為非真實帳號模型辨識重要特徵。
綜合上述，面對社群媒體平台日益複雜的非真實帳號問題，本研究在非真實帳號辨識上提供更具多維的視角，強調多模態分析之應用和貼文行為特徵用於非真實帳號辨識的重要性；此外，本研究所提出之非真實帳號辨識模型在實驗中也展現了顯著的時間效能和高準確性。

關鍵字

社群媒體 ； 機器學習 ； 非真實帳號 ； 虛假帳號 ； 集成學習

並列摘要

Social media plays a crucial role in contemporary information dissemination, with people increasingly relying on it for information access. Consequently, social media has become a battleground for the exposure and reputation-building of enterprises and brands. However, in the post-social media era, the issue of “inauthentic account” has emerged, often utilized for disseminating misleading behavior or statements, spreading viruses, and engaging in fraudulent activities, posing severe risks to society. Notably, in Taiwan, where over seventy percent of social media users predominantly use Facebook, the exploration of inauthentic account identification within the Facebook community becomes imperative.
Recent research on detecting inauthentic account in social media has evolved from early reliance on single or specific data features to incorporating methods that integrate multiple features for identification. Moreover, studies have begun adopting different modal analyses to increase the dimensionality of features used for inauthentic account detection, thereby enhancing accuracy. Nevertheless, systematic review studies indicate a predominant focus on inauthentic account detection research on the Twitter platform, with a comparatively lower emphasis on Facebook, despite its predominant usage among Taiwan’s social media users. Domestic research in this regard remains sparse. This study addresses the gap by reviewing various features dispersed in different literature, collecting data encompassing Facebook user profile information, post data, and extracting over forty features through feature engineering methods. Application of machine learning models, including Random Forest, Support Vector Machine, and Stacking Ensemble Learning, facilitates predictive modeling.
The study's experiments revealed the following findings: Firstly, the inauthentic account identification model employing Random Forest and Ensemble Learning algorithms achieved a recognition accuracy rate of 99%, demonstrating significant practical application value. Secondly, the study found that integrating social interaction features into the inauthentic account identification significantly enhances the model's accuracy. Thirdly, through exploratory analysis, it was discovered that features such as the disparity between image recognition and actual information, average comment counts on posts, and the ratio of check-ins and tags in posts are essential indicators for the inauthentic account identification model.
In conclusion, faced with the increasingly complex issue of inauthentic account on social media platforms, this study aims to provide a comprehensive perspective on inauthentic account identification, emphasizing the application of multimodal analysis and the significance of post behavior features in inauthentic account recognition. Furthermore, the inauthentic account recognition model proposed in this study has also exhibited significant time efficiency and high accuracy in experimental applications.

並列關鍵字

social media ； machine learning ； inauthentic account ； fake account ； ensemble learning

參考文獻

中文文獻：

杜謙和（2017）。一個偵測社群網路中非真實帳號的研究。﹝碩士論文。國立中興大學﹞臺灣博碩士論文知識加值系統。 https://hdl.handle.net/11296/ad2x45。

黃文怡（2020）。基於行為與文本分析的網路粉絲與黑粉人群發現。﹝碩士論文。中原大學﹞臺灣博碩士論文知識加值系統。 https://hdl.handle.net/11296/45eq4n。

楊進倫（2020）。PTT異常帳號偵測。﹝碩士論文。國立臺灣大學﹞臺灣博碩士論文知識加值系統。 https://hdl.handle.net/11296/n9fhs3。

楊雅棠、彭美珍、楊玉玟（2016）。疾病管制署「1922防疫達人」臉書粉絲團營運成效分析。疫情報導，32(15)，312-323。https://doi.org/10.6524/EB.20160809.32(15).001