PrivateRecSys
Privacy-Friendly Recommendation System
The use of recommender systems has grown significantly in recent years, with users receiving personalised recommendations ranging from products to buy, news to read, movies to watch, people to follow. At the same time, recommender systems have become extremely effective revenue drivers for online business. However, producing personalised recommendations requires collecting of users’ data, which makes conventional recommenders effective at the cost of users' privacy. The PrivacyRecSys project aims to develop an open-source toolkit for delivering accurate recommendations while respecting users' privacy. The toolkit will consist of novel privacy-preserving recommender approaches, which modify the state-of-the-art recommender approaches by applying the principles of differential privacy, homomorphic encryption and federated learning.
- The project's own website: https://github.com/privateRecsys/privaterecsys
Why does this actually matter to end users?
Search and discovery is one of the most important and essential use cases of the internet. When you are in school and need to give a presentation, when you are looking for a job, trying to promote your business or finding relevant commercial or public services you need, most of the time you will turn to the internet and more importantly the search bar in your browser to find answers. Searching information and making sure your name, company or idea can be discovered is crucial for users, but they actually have little control over this. Search engines set the terms for what results you see, how your website can be discovered and what information is logged about your searches. What terms are set remains obscure for users and they can only follow the rules laid out for them, instead of deciding on their own what, where and how to find the information they are looking for.
Online search basically is a black box: you enter your question and get an answer, or optimize your site to to end up in the top ten results, but no one has actual control over how it all works. Not only does this make us dependent on search providers, it can (and does) jeopardize your privacy, from the actual query itself to all sorts of sensitive metadata you might leak (other sites you visited, your IP address, other online accounts, etcetera).
So how do we regain control over how we search online? One way to do this is to build transparent, user-centric and privacy-friendly alternatives to popular search solutions. That is what this project aims to do for recommender systems like you find on the bottom part of most webshops (think of 'customers have also bought'). These systems are very lucrative for online businesses, but usually take in a lot of personal data to provide accurate recommendations. This project will show that personalized search and discovery does not have to come at the cost of your privacy by making an open source toolkit for privacy-preserving recommender systems. This way websites and web shops do not have to choose privacy over functionality, but instead combine the two into a more user-friendly online space for everyone.
This project was funded through the NGI0 Discovery Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 825322.
