A combined model of clustering and classification methods for preserving privacy in social networks against inference and neighborhood attacks

Abstract

In the last decade online social networks has gained remarkable attention. Facebook or Google+, are example social network services which allow people to create online profiles and share personal information with their friends. These networks publish details about users while some of the information revealed inside is private. In order to address privacy concerns, many social networks allow users to hide their private or sensitive information in their profiles from the public. In this paper, we focus on the problem of information revelation in online social networks by preserving the privacy of sensitive information in their data using machine learning and data mining algorithms. We show how an adversary can launch an inference or neighborhood attack to exploit an online social network using released data and structure of the network to predict the private information and attributes of users. For this purpose, we propose a new data mining based model that uses neighborhood information and attributes details of a user to infer private attributes of user profiles. The proposed model consists of two main parts: A clustering approach to ensure the k-anonymity and a classification algorithm to preserve the privacy against inference attacks. Finally we explore the effectiveness of some sanitization techniques that can be used to combat such inference attacks, and we show experimentally the success of different neighborhood re-identification strategies. Our experimental results reveal that using combination of data mining algorithm can notably help to preserve private and sensitive information in social network data.