Home Publications Teaching Dataset Press
Fake Account DetectionOnline social systems (e.g., Twitter, Facebook, and Google+) and content ratings services (e.g., Yelp, Google Play, and YouTube) are vulnerable to Sybil attacks, where attackers register a large number of fake accounts to manipulate the systems. For instance, reports from Auguest, 2014 showed that 8.5% of Twitter active users were fake/Sybils; and it was reported in August, 2012 that 9.2% of Facebook users were fake/Sybils. These abusive accounts leverage their access to millions of benign users to disseminate scams, carry out phishing attacks, distribute malware, and harvest private user data. On Yelp, an attacker (e.g., the owner of a restaurant) could register many accounts and write fake reviews to manipulate the reputation of any restaurant.
Our work aims to detect Sybil accounts using the social relationships between users. The intuition is that it is hard for attackers to establish trust relationships between Sybil accounts and benign users, though they can manipulate the Sybil accounts arbitrarily. Previous work leveraged either known benign users or known Sybil accounts, but not both. From machine learning perspective, they are outlier detection approaches. We designed a scalable semi-supervised learning framework called SybilBelief, which leverages pairwise Markov Random Fields and Loopy Belief Propagation and can incorporate both known benign users and known Sybil accounts. Our ongoing research applies SybilBelief to detect Sybil accounts in a large-scale Twitter network dataset consisting of 42 million nodes and 1.5 billion edges.