The use of trust for onion routing has been proved effective in thwarting malicious onion routers. However, even state-of-the-art trust-based onion routing protocols still suffer from two key limitations in protecting anonymity. First, these protocols have no means to verify the correctness of the trust they rely on. Second, they run a high risk of being deanonymized by an inference attack due to biased trust distributions. In this paper, we propose SGor, a trus tg raph based onion routing that mitigates the key limitations of trust in protecting anonymity. SGor is novel with three unique properties. First, SGor aggregates group trust from mutual friends to verify the correctness of users’ trust assignments. Second, SGor employs an adaptive trust propagation algorithm to derive global trust from trust graph. The global trust removes the restriction of users’ local knowledge and defeats inference attacks by guiding users to discover and trust more honest routers (i.e., reducing the bias of trust distribution). Third, SGor is designed to operate in a fully decentralized manner. This decentralized design mitigates the leakage of a priori trust relationships. We evaluate SGor with simulation-based experiments using several real-world social trust datasets. The experimental results confirm that SGor can mitigate key limitations in the use of trust for protecting anonymity but introduces only a few overheads.