Deriving private numeric data in privacy-preserving collaborative filtering systems

There are a great number of methods introduced to offer reliable and accurate recommendations in privacy-preserving collaborative filtering systems without disclosing individual or data holders' private data. Privacy has become a fundamental principle in privacy-preserving collaborative filtering systems. Several data disguising methods have been proposed to alleviate users' privacy concerns. Randomization, which is one of the such proposed methods, is commonly utilized to protect individual private data while allowing collaborative filtering systems to produce decent recommendations. However, recent studies show that it is likely to reconstruct private data from the data disguised by randomization. The dissertation scrutinizes whether the private numeric data can be reconstructed from central, partitioned and distributed data-based privacy-preserving schemes by planning privacy attack scenarios. In central data-based schemes, noise from z-scores must be eliminated first by employing the noise elimination methods. A new method is proposed to derive the original ratings from the estimated z-score values utilizing public data. It is also crucial to reconstruct rated items when users inconsistently perturb their confidential data. In partitioned data based-schemes, deriving other party's private data from the aggregate results are investigated. In distributed data-based schemes, data holders might coalesce against one holder to derive its private data. In addition, solutions are proposed by utilizing various domain-related auxiliary information and characteristic features of collaborative filtering systems to enhance the reconstructions. Theoretical analysis and experimental outcomes demonstrate that the proposed approaches help attackers derive a meaningful amount of private data in certain circumstances.