Machine Learning based Stochastic Techniques for Collaborative Privacy in Social Recommender Services

Mohamed Khamis Elmesiry, Ahmed (2014) Machine Learning based Stochastic Techniques for Collaborative Privacy in Social Recommender Services. PhD thesis, Waterford Institute of Technology.

[thumbnail of Elmesiry_Ahmed_FullThesis.pdf]

Download (90MB) | Preview


A variety of online social services have been developed over the last decade. They have all had a profound effect on today’s society. With the emergence of Web 2.0 and the popularity of social media, there has been a growing demand to provide services supporting social network platforms. New services are constantly being develpoed, where an increasing volume of personal data is being processed in return for personally tailored services. This result in creating mature systems to satisfy users’ needs is referred to as social recommender services that create an inevitable trend driven by mutual benefits. The penetration of these services has been relatively slow recently since there still exists different viewpoints regarding the exploitation (and thus potential) of those services amongst researchers and users. One major concern regarding their adoption lies in privacy considerations of the users while using these services. With the increasing amount of personal data that users distribute over different services, there is an increased probability for identity fraud, profiling and linkability attacks, that not only poses a threat to those people’s personal dignity and affects different aspects of their lives, but also to societies as a whole. In most cases, this can prevent users from fully embracing these services whereas most of the “privacy-concerned” systems that have been developed so far, are either based on a trusted third-party model or on some generalized architecture. This thesis focuses primarily on social recommender services which are of great interest. On the one hand, they lay the groundwork for new innovative applications but on the other hand they pose numerous unique challenges to privacy. We studied the privacy problem faced by people in sharing their profiles’ preferences within various scenarios of social recommender services. We proposed and developed a collaborative privacy approach for preserving users’ profile privacy and we have applied this approach to representative scenarios: (I.) Recommender service for IPTV content providers; (II.) Data Mash-up services for IPTV recommender services and (III.) Community discovery and recommendation services for implicit social groups (conference organization and university campus). Location based recommendation services, mobile jukebox content recommender services and pervasive healthcare services were studied and enhanced as well in order to show the applicability of our approach. We discussed how our approach could handle the privacy problem in these scenarios. In Addition, the proposed collaborative privacy framework was developed as a middleware that hosts a set of components to execute a two stage concealment process with novel stochastic techniques. Each stage in the two stage concealment process is carried out by completely different parties depending on their role in the coalition. The proposed middleware as well as the set of components and techniques that is employed in its implementation, permit the end-users to control the privacy of their released data while interacting with social recommender services. This kind of approach is quite flexible and can easily be adopted in conventional social recommender services because it is executed on the user side and takes advantage of the social structure that is offered by the online social service without the need for significant modifications at the service provider side. The attained accuracy and privacy levels for the data concealed using the proposed stochastic techniques in different scenarios were evaluated. Moreover, attacks on such concealed data were presented to demonstrate the stability of our proposed techniques against such attacks. Finally, we applied off-the-shelf recommendation techniques to make referrals as a show case. Therefore, the experimental results show that the proposed approach obtains accurate results similar to unsecured services, while at the same time meeting users’ privacy concerns.

Item Type: Thesis (PhD)
Uncontrolled Keywords: Machine Learning based Stochastic Techniques, Social Recommender Services
Departments or Groups: *NONE OF THESE*
Divisions: School of Science > Department of Computing, Maths and Physics
Depositing User: Derek Langford
Date Deposited: 02 Oct 2014 16:02
Last Modified: 22 Aug 2016 10:27

Actions (login required)

View Item View Item