Document Type |
: |
Thesis |
Document Title |
: |
DETECTING MALICIOUS CAMPAIGNS IN SOCIAL NETWORKS BASED ON A COLLECTIVE BEHAVIOR ANALYSIS الكشف عن الجماعات الخبيثة في شبكات التواصل الاجتماعي بناء على تحليل السلوك الجماعي |
Subject |
: |
Faculty of Computing and Information Technology |
Document Language |
: |
Arabic |
Abstract |
: |
Social spam involves several malicious activities threatening social network users by sharing unwanted, misleading, or spiteful content to either send unwanted ads, manipulate public opinion, or spread harmful malware. This has led to the development of considerable researches and approaches to combat such practices in the social network. Most works, however, were developed to address the detection of English language spammers with little efforts for the Arabic language spammers. This thesis, therefore, first provides a comprehensive analysis of the characteristics of malicious Arab campaigns on Twitter. Besides exposing their spamming tactics, these accounts were found to be more successful in avoiding Twitter suspension than previously reported spammers in the literature. Based on the outcomes of the analysis, two semi-supervised algorithms were adopted, and a set of 16 features were used to identify the campaigns’ accounts automatically. The proposed model performance was tested on a dataset of 1685 accounts, and the results show that the model achieved a 0.91 accuracy. Since there is a considerable need for real-time detection techniques to filter malicious or low-quality tweets, we have also developed a real-time deep-learning approach that utilizes tweet textual data to detect spam content and spammers profiles. The proposed approach was evaluated in a real-world dataset that includes a wide range of low-quality tweets, and our model demonstrates superior performance compared to existing solutions in both accuracy and F1 measure (0.98). We also proposed a lightweight deep learning approach to classify Twitter accounts as spam or genuine accounts based solely on the account recent tweets. The lightweight method yielded high accuracy scores and an F1 measure (0.98). In less than five milliseconds, our real-time approach was able to classify a tweet and an account in less than a second. |
Supervisor |
: |
Dr. Kawthar Moria |
Thesis Type |
: |
Master Thesis |
Publishing Year |
: |
1441 AH
2020 AD |
Co-Supervisor |
: |
Dr. Areej Alhothali |
Added Date |
: |
Wednesday, June 3, 2020 |
|
Researchers
ريم مسفر الحارثي | Alharthy, Reem Mesfer | Researcher | Master | |
|