Crowdguard: Characterization and early detection of collective content polluters in online social networks

Recently, content polluters post malicious information in Online Social Networks (OSNs), which is a more and more serious problem that poses a serious threat to the privacy information, account security, user experience, etc. They continuously simulate the behaviors of legitimate accounts in various ways, and evade detection systems against them. In this paper, we focus on one kind of content polluter, namely collective content polluter (hereinafter referred to as CCP). Existing works either focus on individual polluters or require long periods of data records for detection, making their detection methods less robust and lagging behind. It is thus necessary to analyze the characteristics of collective content polluters and study the methods for early detection. This paper proposes a CCP early detection method called CrowdGuard. It analyzes the crowd behaviors of collective content polluters and legitimate accounts, extracts distinctive features, and leverages the Gaussian Mixture Model (GMM) method to cluster the two groups of accounts (legitimate users and polluters) to achieve early detection. Using the public dataset including thousands of collective content polluters on Twitter about a political election, we design an experimental scenario simulating early detection and evaluate the performance of CrowdGuard. The results show that CrowdGuard outperforms existing methods and is adequate for early detection.