Application of differential privacy to recommendation systems

Abstract

The deluge of data on the web, online products, and services has made recommendation systems an integral part of the Web realm. They are employed in a wide range of applications starting from eCommerce sites, through the Social Web, to healthcare apps. Recommenders are leveraged to enhance the product sales, to help user in quick decision making, and to suggest relevant products to users from massive product catalogue. Typically, recommender systems rely on users personal information to train the system, to prioritize the relevant items to a specific user based on the previous preferences and to predict the rating for new products based on the user s behavior. However, such data usage in recommender systems hampers the user s sensitive information and poses a severe threat to individual privacy. newlineThis thesis addresses the problems of privacy preservation in recommender systems using differential privacy. Differential privacy features a provable privacy guarantee, but challenges the application of the same in the recommender systems. The following are some of the challenges in applying differential privacy to recommender: The dataset s sparsity aids attackers to launch various attacks on the recommenders, newlineHigh dimensionality of data in recommenders in turn results in large noise addition when perturbation based techniques are used, Maintaining a balance between privacy and utility in perturbation based techniques. newline newline newline