High Performance Private Recommender Systems
In current recommender systems user rating data is linked to specific users. If the server is compromised or the service provider chooses to share this data (even if "anonymised") then that data can release unexpected information, e.g. by combining IMDB review data with the anonymised netflix ratings users were deanonymised. In BLC users submit ratings using group identities, known as nyms. Each group contains many thousands of users and so it is v hard to guess which ratings were submitted by any specific user. This provides a strong "hiding in the crowd" type of privacy. Importantly, it also gives state of the art recommendation performance i.e. there is no trade-off between accuracy and privacy here.
Summary of BLC recommender system.
- BLC: Private Matrix Factorization Recommenders via Automatic Group Learning.Alessandro Checco, Giuseppe Bianchi, Douglas J. Leith, Tech Report Oct 2015
- Recommending Access Points to Individual Mobile Users via Automatic Group Learning. Bahar Partov, Douglas J. Leith, Alessandro Checco. Tech Report Oct 2016
Code
Send me an email if you'd like python or matlab code implementing the BLC recommender.
Private Search
- Don't let Google know I'm lonely!. Pol Mac Aonghusa, Douglas J. Leith, ACM Trans Security and Privacy, 2016
- It wasnt me! Plausible Deniability in Web Search. Pol Mac Aonghusa, Douglas J. Leith, Tech Report Sept 2016