Developing techniques for robustness of privacy-preserving distributed collaborative filtering

Success of collaborative filtering systems strongly depend on having adequate data. Due to customers' shopping habits and increasing number of e-commerce sites, data collected for referral purposes might be distributed among various sites. Therefore, especially for newly established companies, offering recommendation services might turn out to be a trouble, due to lack of qualified data. To overcome this challenge, collaboration of online vendors on distributed data while preserving privacy has become an important topic. Researchers have proposed several privacy-preserving distributed collaborative filtering schemes, which enable collaboration of online vendors, even the competing ones, on distributed data without jeopardizing privacy. However, such schemes have not been evaluated in terms of robustness against attacks. If manipulating the outcomes of privacy-preserving distributed collaborative filtering algorithms by injecting fake profiles is possible, shilling attacks might be an obstacle for collaboration. Online vendors, who are unsure of being subject to shilling attacks, might refrain from cooperation, even if they need it for offering more useful recommendation services to their customers. In this dissertation, robustness of state-of-the-art privacy-preserving distributed collaborative filtering schemes proposed for arbitrarily distributed data are analyzed against shilling attacks. A new attack strategy that can be applied on arbitrarily distributed data, and used in generation of distributed adaptations of formerly proposed attack models is outlined. Empirical studies show that attacks generated by the proposed strategy are effective in manipulating predicted outcomes, hence, despite privacy, these schemes are not resistant to attacks. The reasons of why existing shilling attack detection methods cannot be directly employed on arbitrarily distributed data are discussed. To protect these algorithms against attacks, distributed version of a well-known classification-based attack detection method is proposed, which can operate on arbitrarily distributed data. Real data-based experiments demonstrate that the proposed detection method is able to identify distributed attack profiles on arbitrary data with privacy. Moreover, the need for collaboration in detection of distributed attacks is exposed with experimental analyzes.