Online Social Networks (OSN) are increasingly becoming victims of Sybil attacks. These attacks involve creation of multiple colluding fake accounts (called Sybils) with the goal of compromising the trust underpinnings of the OSN, in turn, leading to security and the privacy violations. Existing mechanisms to detect Sybils are based either on analyzing user attributes and activities, which are often incomplete or inaccurate or raise privacy concerns, or on analyzing the topological structures of the OSN. Two major assumptions that the latter category of works make, namely, that the OSN can be partitioned into a Sybil and a non-Sybil region and that the so-called “attack edges” between Sybil nodes and non-Sybil nodes are only a handful, often do not hold in real life scenarios. Consequently, when attackers engineer Sybils to behave like real user accounts, these mechanisms perform poorly. In this work, we propose SybilRadar, a robust Sybil detection framework based on graph-based structural properties of an OSN that does not rely on the traditional non-realistic assumptions that similar structure-based frameworks make. We run SybilRadar on both synthetic as well as real-world OSN data. Our results demonstrate that SybilRadar has very high detection rate even when the network is not fast mixing and the so-called “attack edges” between Sybils and non-Sybils are in the tens of thousands.