Sharing images on social network sites has become a part of daily routine for more and more online users. However, in face of the considerable amount of images shared online, it is not a trivial task for a person to manually configure proper privacy settings for each of the images that he/she uploaded. The lack of proper privacy protection during image sharing could raise many potential privacy breaches of people's private lives that they are not aware of. In this work, we propose a privacy setting recommender system to help people effortlessly set up the privacy settings for their online images. The key idea is developed based on our finding that there are certain correlations between a number of generic patterns of image privacy settings and image tags, regardless of the image owners' individual privacy bias and levels of awareness. We propose a multi-pronged mechanism that carefully analyzes tags' semantics and co-presence to derive a set of suitable privacy settings for a newly uploaded image. Our system is also capable of dealing with cold-start problem when there are very few image tags available. We have conducted extensive experimental studies and the results demonstrate the effectiveness of our approach in terms of the policy recommendation accuracy.