Collaborative and recommendation-based computer systems are often plagued by attackers who create fake or malicious identities to gain more influence in the system — such attacks are often referred to as “Sybil attacks”. We propose a new statistical model and associated learning algorithms for detecting Sybil attacks in a collaborative network, called the latent community (LC) model. The LC model is hierarchical, and groups the nodes in a network into closely linked communities that are linked relatively loosely with the rest of the graph. Since the author of a Sybil attack will typically create many false identities and link them together in an attempt to gain influence in the network, a Sybil attack will often correspond to a learned community in the LC model. Evaluation of the LC model using real-world networks validates the model and shows that it can be superior to competitive algorithms from network security literature for detecting Sybil attacks.