Balancing Personalisation and Privacy: User Matching and Recommender Systems in Social Networks

Online social networks enable people to connect and share information in the virtual space. These platforms offer a wide range of features such as user matching for friend discovery, content sharing and recommendations for personalised experience and so on. User matching and content recommendations are crucial components for online social networks, as user matching allows individuals to establish a network of connections with other users on the same platform, while content sharing and recommendations enhance user engagement and experience. However, both user matching and recommendations leverage user data for insights, such as user profiles for friend discovery and past activities such as ratings and comments for tailored recommendations. This raises privacy concerns about the confidentiality of the data, given its sensitive nature. Cyber incidents such as data breaches have occurred on social networks, revealing private user data to the public, which affects not only the platform but also the individuals whose data is compromised. While previous papers devoted to privacy-preserving user matching and recommender systems have aimed to mitigate privacy concerns, such as applying cryptography, many of them lack practicability or functionality due to high computational overhead and limitations imposed by the underlying privacy-preserving mechanisms. This thesis focuses on designing practical privacy-preserving solutions for user matching and recommender systems in online social networks, enabling personalised experiences while protecting user data. Considering the lack of practicability in existing solutions, we focus on mitigating the performance issues with these systems, while upholding user privacy. We propose a privacy-preserving user matching system using a hardware-based security method that enables trusted execution and ensures robust security in cloud computation. Subsequently, we introduce efficient privacy-preserving recommender systems utilising cryptographic techniques to protect user privacy, and incorporating a secure clustering to improve their practicability. The proposed privacy-preserving recommender systems are further enhanced by taking the user matching system into consideration, as well as exploiting advanced optimisation techniques to further reduce both computational and communication overheads. Lastly, we propose an efficient and privacy-preserving recommender system based on machine learning. A more lightweight privacy-preserving mechanism is applied to the system, allowing distributed learning without compromising data privacy. Comprehensive evaluations of our proposed systems show promising results and outperform existing works in terms of performance. Together, we provide privacy-preserving systems to empower social networks to offer user matching and recommender systems while safeguarding user privacy.