Author:
Lixin Fu
Affiliation:
University of North Carolina at Greensboro, United States
Keyword(s):
Sybil Detection, Spam Detection, Social Networks.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Data Mining
;
Databases and Information Systems Integration
;
Enterprise Information Systems
;
Sensor Networks
;
Signal Processing
;
Society, e-Business and e-Government
;
Soft Computing
;
Software Agents and Internet Computing
;
Web 2.0 and Social Networking Controls
;
Web Information Systems and Technologies
Abstract:
Social media becomes a common platform for millions of people to communicate with one another online. However, some accounts and computer generated robots can greatly disrupt the normal communications. For example, the fake accounts can simultaneously "like" or "dislike" a tweet, therefore, distort the true nature of the attitudes of real human beings. They collectively respond with similar or the same automate messages to influence sentiment towards certain subject or a tweet. They may also generate large amounts of unwanted spam messages including the irrelevant advertisements of products and services. Even worse, some messages contain harmful phishing links that steal people's sensitive information. We propose a new system that can detect these disruptive behaviours on OSNs. Our methods is to integrate several sybil detection models into one prediction model based on the account profiles, social graph characteristics, comment content, and user feedback reports. Specifically, we
give two new detection algorithms that have better prediction accuracy than that of the state--of-the-art systems and real time performance. In addition, a prototype system including the software modules and real and synthetic data sets on which comprehensive experiments may confirm our hypothesis. Currently most sybil detection algorithms are based on the structural connections such as few connections of densely connected Sybil communities to normal nodes. Their detection accuracy is mixed and not well. Some algorithms are based on machine learning. The different approaches are separated. We expect our new model will more accurately detect the disruptive behaviour of fake identities with high positive rates and low false negative rates.
(More)