A Robust Privacy-Preserving Recommendation Algorithm

Privacy-preserving collaborative filtering schemes are key recommender system technologies for e-commerce field. They focus on alleviating information overload problem by providing personalized recommendations without deeply jeopardizing customers’ privacy. Like their non-private versions, privacy-preserving recommendation methods might be easily subjected to profile injection attacks for manipulating produced recommendations in favor or disfavor of certain products. Clustering-based prediction schemes have shown to be effective in distinguishing bogus profiles from genuine ones; and they are relatively more resilient than memory-based methods against shilling attacks in non-private environments. Motivating from this fact, we propose bisecting k-means clustering-based privacy-preserving recommendation algorithm as a robust recommendation algorithm, which was formerly proposed as a scalable and accurate private recommendation scheme, against previously defined several private shilling attack strategies. We then investigate the algorithm with respect to robustness. Thus, we perform some real data-based experiments on a benchmark data set using six profile injection attacks. We empirically show that the algorithm performs in a robust manner with insignificant alterations in predicted values when fake profiles are injected. The reason for this phenomenon is that the algorithm is able to cluster fake profiles together.