Nowadays, the proliferation of handheld devices embedded with multiple mobile sensors and the growth of fast communication and data processing technologies have contributed to the emergence of a wide variety of online services, including location-based services. These services facilitate users' daily lives with a broad range of applications that offer users personalized and customized information about their surroundings according to their location. While these services have undeniably become essential and indispensable to our society today and especially in the future, it is necessary to underline and understand the risks and threats affecting users. Indeed, large amounts of mobility data are being gathered, stored, and processed by service providers or third parties without necessarily the users' consent. As a consequence, users' privacy is threatened, and thus many sensitive information, such as the user's identity, home or workplace address, or even religious beliefs or health status can be inferred and leaked. In this context, it becomes urgent to devise mechanisms that allow users to securely and safely access location-based services without disclosing their private lives.

To address this challenge, many efforts aim to enhance privacy by proposing new location privacy protection mechanisms (LPPMs). These efforts are not only motivated by the research community, but authorities and organizations increasingly establish new laws and regulations to reframe the collection, storage, and manipulation of users' mobility data. In this direction, location privacy risk assessment (LPRA) is designed to assess the privacy risks of sharing mobility data to raise users' awareness about their privacy. In this manuscript, we use the re-identification risk, which aims at re-linking an anonymous mobility data to its originating user as a means of LPRA.

In this thesis, we first propose MOOD, a centralized user-centric protection system that aims to protect the mobility data of all users and in particular those who are not protected by any individual LPPM. MOOD uses the composition of several LPPMs and incorporates the re-identification risk assessment before publishing the protected data. However, it requires a trusted proxy server to perform both the obfuscation process and the re-identification risk assessment. Although existing protection methods aim to eliminate the trusted proxy server, the privacy risk assessment still needs to centralize the mobility data.
That is why we propose SAFER, a novel privacy risk assessment metric, developed on the user side, to estimate how unique a user's mobility data is among a group of participating users. SAFER follows a federated learning approach to build a global knowledge without accessing raw users' mobility data in a central entity. Finally, we propose EDEN, a user-side mobility data protection system that automatically selects the best LPPM and its corresponding configuration that resists the re-identification risk assessment without sending the raw mobility data outside the user's device thanks to the federated learning paradigm