Recommender Systems on Homomorphically Encrypted Databases for Enhanced User Privacy

Detailed customer statistics play an important role for merchants to distribute targeted advertising and derive sales statistics. For that, merchants are strongly interested in correlations between product groups to generate proper recommendation models. Systems that collect customer data are usually profile-based and pose a threat to customer privacy. For further improving their advertising cooperating merchants can share their customer profiles. We propose to employ homomorphic encryption to ensure all customer profile data remains protected without revealing the merchants operational secrets. Only a set of predefined operations can be performed on aggregated records of these encrypted data to calculate the desired correlations and still protect privacy. We perform an evaluation on the efficiency and feasibility of our prototypical implementation using simulated data. Our experiments show, that such a system can be implemented using commercial off-the-shelf hardware and software with negligible additional effort.