Binary data reconstruction in privacy-preserving recommendation algorithms

Collaborative filtering systems have become very popular with the frequent use of the Internet to offer reliable recommendations to users. Ratings for such systems could be in a binary or numeric scale, and data supplied by users could be stored in a central-server, distributed among two- or multi-party or even peers could come together for collaborative filtering purposes. Collaborative filtering systems rely on true user feedbacks in order to produce accurate recommendations. However, users of such systems might be reluctant to provide their true opinions if they feel that their confidential data might be used other than the initial purpose of data collection. Such resistances to participate in collaborative filtering systems might hamper the recommendation quality. At this point, privacy-preserving collaborating filtering systems take the privacy concerns into the primary consideration without sacrificing the recommendation quality. Therefore, users are convinced to provide their true inputs as well as receive quality recommendations by the measures taken by privacy-preserving collaborative filtering systems. However, these measures should be investigated if the claimed privacy-preservation is really maintained. The objective of this dissertation is to derive the original binary ratings, which are promised to be preserved, from the perturbed binary ratings in different privacy-preservation protocols under different data partitioning scenarios including central server-based, distributed between two- and multi-party and peer-to-peer collaboration. Auxiliary information is utilized throughout the dissertation to improve the reconstruction accuracy or circumvent the bottlenecks to derive the original ratings due to the privacy-preservation protocols.