Photo privacy has raised a growing concern with the advancements of image analytics, face recognition, and deep learning techniques widely applied on social media. If properly deployed, these powerful techniques can in turn assist people in enhancing their online privacy. One possible approach is to build a strong, automatic and dynamic access control mechanism based on analyzing the image content and learning users sharing behavior. This paper presents a model for context-dependent and privacy-aware photo sharing based on machine learning. The proposed model utilizes image semantics and requester contextual information to decide whether or not to share a particular picture with a specific requester at certain context, and if yes, at which granularity. To evaluate the proposed model, we conducted a user study on 23 subjects and collected a dataset containing 1’018 manually annotated images with 12’216 personalized contextual sharing decisions. Evaluation experiments were performed and the results show a promising performance of the proposed model for photo sharing decision making. Furthermore, the influences of different types of features on decision making have been investigated, the results of which validate the usefulness of pre-defined features and imply a significant variance between users sharing behaviors and privacy attitudes.