Inference of high-level context is becoming crucial in development of context-aware applications. An example is social context inference - i.e., deriving social relations based upon the user's daily communication with other people. The efficiency of this mechanism mainly depends on the method(s) used to draw inferences based on existing evidence and sample information, such as a training data. Our approach uses rule-based data mining, Bayesian network inference, and user feedback to compute the probabilities of another user being in the specific social relationship with a user whose daily communication is logged by a mobile phone. In addition, a privacy mechanism is required to ensure the user's personal integrity and privacy when sharing this user's sensitive context data. Therefore, the derived social relations are used to define a user's policies for context access control, which grant the restricted context information scope depending on the user's current context. Finally, we propose a distributed architecture capable of managing this context information based upon these context access policies.