De-anonymizing Social Networks User via Profile Similarity

Abstract

Nowadays people are likely to reveal different aspects of life on different websites. Given a user's account information on one site, we can find his or her other accounts on other websites with the help of profile matching, which can benefit multiple application domains, including recommendation, privacy, security and so on. Traditional profile matching method makes use of as many attributes as possible and calculates the similarity of attribute values one by one, which is not fit for the realistic scenario with large-scale data and lack of common attributes. Therefore, in this paper, we proposed a new profile matching method. We extracted information about user identities on about.me, Twitter and Github; We used real name, username, location, and url to generate the feature of the user. We applied several similarity measures (e.g. Jaro Winkler similarity, Levenshtein distance and N-gram distance) to measure the similarity of the user profiles on different websites. Using the most promising similarity measure and parameters, we achieved a high reliability with a recall over 95% for a 98% precision, similar to the basic method and greatly reduced the running time.