Many online services, such as recommender systems, email, and social networks collect user data, which is then used for both personalization and monetization. Although the latter enables services to be free, recent upsurge in privacy violations makes users realize that these services come at a hidden cost of potentially exposing their private data. As the loss of privacy in the digital age becomes more rampant, there is a clear need to gain back user trust by designing services that protect user privacy. In this talk I will show how, using a combination of homomorphic encryption and garbled circuits, we introduced strong privacy guarantees to two big-data analysis algorithms, namely, linear regression and matrix factorization. I will present a system that learns a linear model without learning anything about the private input data other than the output model, and a recommender system that can recommend items to users without learning the item-ratings submitted by the users. Unlike previous efforts, the systems I will present can operate on truly large datasets in practical time, memory and computation constrains.