Towards Privacy-preserving Users’ Data via Attribute Resolutions in Context-Aware Recommender Systems

Lobbezoo, Micha

Towards Privacy-preserving Users’ Data via Attribute Resolutions in Context-Aware Recommender Systems

Files

Lobbezoo, M. s-1019806-BSc-Thesis-2022.pdf (390.94 KB)

Authors

Lobbezoo, Micha

Issue Date

2022-06-26

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/15969

Abstract

Anonymised data sets are published for machine learning challenges, where participants can attempt to improve the performance of current models. Attacks such as joining attacks showed that such data sets are not always as anonymised as assumed, which results in leaks of user data. Another risk with such data sets is the inference of users’ private attributes, which can be implicitly stored in the data. In this thesis, we will focus on privacy preserving the users’ private attributes in recommender systems. Possible attacks on the users privacy are discussed and a new approach is proposed, which is attribute resolution. This method uses intervals instead of the more exact value of the user attribute, which decreases the resolution of this attribute. With an inference attack, the leakage of the users’ personal attribute age from the user-item matrix is demonstrated. Then we use the RankFM recommender system to generate and evaluate recommendations with and without the age of the user as user side information and compare the performance of the recommendations, showing that the performance does improve when the model knows the age of the user. Lastly, we run the recommendation algorithm with different sets of intervals of the age attribute as side information to see which effect attribute resolution has on the recommendation performance. Here we can see a minimal increase in recommendation performance when more specific user side information is available to the system. Overall, attribute resolution could be a solution to preserve the users’ privacy.