[1711.07638] Towards a More Reliable Privacy-preserving Recommender System
We plan to extend this framework to content-based models and other models that also use gradient descent to learn latent representations (e.g., deep learning framework) so the applications will not be limited to only recommender systems.

Abstract: This paper proposes a privacy-preserving distributed recommendation
framework, Secure Distributed Collaborative Filtering (SDCF), to preserve the
privacy of value, model and existence altogether. That says, not only the
ratings from the users to the items, but also the existence of the ratings as
well as the learned recommendation model are kept private in our framework. Our
solution relies on a distributed client-server architecture and a two-stage
Randomized Response algorithm, along with an implementation on the popular
recommendation model, Matrix Factorization (MF). We further prove SDCF to meet
the guarantee of Differential Privacy so that clients are allowed to specify
arbitrary privacy levels. Experiments conducted on numerical rating prediction
and one-class rating action prediction exhibit that SDCF does not sacrifice too
much accuracy for privacy.

‹Figure 1: Secured distributed architecture using MF as an example. (Introduction)Figure 2: The overview of SDMF framework. (Framework Overview)

$$$$Figure 3: Representing the existence of ratings by a binary matrix (left). Using RR to sample a noised bit vector St i from binary vector Bi of user i (upper right). Two-stage RR algorithm (lower right). (Randomized Response (RR) Algorithm)Figure 4: Steps to sample eij for unrated items. (Number of Gradients Sent to the Server)

$$$$Figure 5: Task 1: Comparison of different g with fixed I . Figure 6: Task 1: Comparison of different I with fixed g. Figure 7: MovieLens-100K (Experimental Results)Figure 8: Task 1: Comparison of different g with fixed I . Figure 9: Task 1: Comparison of different I with fixed g. Figure 10: MovieLens-1M (Experimental Results)

$$$$Figure 11: Task 1: Comparison of different g with fixed I . Figure 12: Task 1: Comparison of different I with fixed g. Figure 13: Netflix (subsampled) (Experimental Results)