Members of the defense committee:
- Prof. Dr. Yves Le Traon, Université du Luxembourg, chairman
- Dr. Qiang Tang, LIST Luxembourg, vice-chairman
- Prof. Dr. Peter Y.A. Ryan, Université du Luxembourg, supervisor
- A-Prof. Dr. Melek Önen, EURECOM France, member
- A-Prof. Dr. Claudia Diaz, KU Leuven, Belgium, member
Abstract:
Cloud services enable companies to outsource data storage and computation. Resource-limited entities could use this pay-per-use model to outsource large-scale computational tasks to a cloud-service-provider. Nonetheless, this on-demand network access raises the issues of security and privacy, which has become a primary concern of recent decades. In this dissertation, we tackle these problems from two perspectives: data confidentiality and result in integrity.
Concerning data confidentiality, we systematically classify the relaxations of the most widely used privacy preserving technique called Differential Privacy. We also establish a partial ordering of strength between these relaxations and enlist whether they satisfy additional desirable properties, such as composition and privacy axioms.
Tackling the problem of confidentiality, we design a Collaborative Learning game, which helps the data holders to determine how to set the privacy parameter based on economic aspects. We also define the Price of Privacy to measure the overall degradation of accuracy resulting from the applied privacy protection. Moreover, we develop a procedure called Self-Division, which bridges the gap between the game and real-world scenarios.
Concerning result integrity, we formulate a Stackelberg game between outsourcer and outsourcee where no absolute correctness is required. We provide the optimal strategies for the players and perform a sensitivity analysis. Furthermore, we extend the game by allowing the outsourcer no to verify and show its Nash Equilibriums.
Regarding integrity verification, we analyze and compare two verification methods for Collaborative Filtering algorithms: the splitting and the auxiliary data approach. We observe that neither methods provide a full solution for the raised problem. Hence, we propose a solution, which besides outperforming these is also applicable to both stages of the algorithms.