Analyzing spammers' social networks for fun and profit

SybilBelief: A Semi-Supervised Learning Approach for Structure-Based Sybil Detection

“…Sybil classification approaches such as SybilLimit [9] and SybilInfer [10] do not incorporate information about known Sybil labels (limiting detection accuracy, as shown in Section VI), are not resilient to label noise 1 , and are not scalable. Sybil ranking approaches such as SybilRank [13] and CIA [14] incorporate information about either known benign or known Sybil labels, but not both. They are also not resilient to label noise.…”

“…Furthermore, we find that our algorithm can detect the incorrect labels with 100% accuracy. Impact of community structure: Most existing Sybil detection mechanisms [7]- [14] rely on the assumption that the benign region is fast mixing. However, Mohaisen [30] showed that many real-world social networks may not be as fastmixing as was previously thought.…”

“…We compare SybilBelief with two classical Sybil classification mechanisms, i.e., SybilLimit [9] and SybilInfer [10], and two recent Sybil ranking mechanisms, i.e., SybilRank [13] and Criminal account Inference Algorithm [14]. Table II summarizes the notations of these algorithms.…”

“…As was proposed in the original paper [14], we set the restart parameter α to 0.85. Note that CIA can only incorporate labeled Sybil nodes.…”

“…However, existing structure-based approaches suffer from one or more of the following drawbacks: (1) they can bootstrap from either only known benign [8]- [10], [12] or known Sybil nodes [14], limiting their detection accuracy (see Section VI), (2) they cannot tolerate noise in their prior knowledge about known benign [13] or Sybil nodes [14], and (3) they are not scalable [7]- [12].…”