Can On-line Social Network Users Trust That What They Designated as Confidential Data Remains So?

Internet users in general and on-line social networks users in particular are becoming more savvy about masking data they consider private. However, some of this masked data may be inferable from other data the user has not masked. Furthermore, even if a user masks all its data, it may still be inferable from the unmasked data of its friends, due to affinities in likes and personal attributes. In contrast to the conventional data mining approach, in which a model is built for all users, we build a rule set which is individualized for each user. In this paper we propose a novel rule induction approach (that incorporates predictive metrics) which enable a user to evaluate the potential risk incurred by unmasked attributes, friends´ attributes and also the risk of befriending new users. We find that all of these risks are quantifiable and a risk ranking of attributes and friends/potential friends can be individualized for each user. We give examples and use cases and confirm the effectiveness of the approach, using a sophisticated synthetic OSN-data to define risk attribute and user combinations which coincide with the risk ranking produced by our algorithm.