Defence Teimour Hosseinalizadeh: "Privacy-aware control of cyber-physical systems: Analysis and synthesis"

Promotors: 1^st promotor: Prof. Nima Monshizadeh Naini, 2^nd promotor: Fatih Turkmen, 3^rd promotor: Prof. Claudio De Persis

Abstract:

Internet connectivity has been extending its reach from servers and clients into embedded systems that connect to industrial processes, vehicles, home automation and so forth. While the resulting cyber-physical systems (CPSs) create an enormous potential for the development of smarter and more efficient manufacturing, traffic and energy management they also introduce the challenge of preserving the privacy for the data of involved systems. Motivated by these two factors, this thesis aims to both analyse and synthesize privacy-preserving schemes for the control of CPSs. In the first part of this thesis, we concentrate on cloud-based computations and the role of system theory-based approaches in achieving privacy. We investigate to what extent Cloud is successful in determining privacy sensitive parameters while solving a model predictive control (MPC) problem by drawing on common side-knowledge. The analysis shows that in separate and dense forms of MPC, algebraic transformations are vulnerable to
side-knowledge. This paves the way for designing an outsourcing scheme for cloud-based data-driven control problem. While constraint by factors such as real-time applications and being justified, we combine transformation-based methods and robust control theory to have a cloud-based data-driven control design. The scheme preserves privacy for open-loop and closed-loop matrices while applicable to system with and without disturbance. The next part of this thesis draws on perturbation-based methods, as strong and universal solutions, and focuses on keeping the initial state of a system private. The motivation originates from many data-releasing systems where crucial signals must be shared with a monitoring center. We show that by using correlated Gaussian noise any
confusion set in the form of an hyper-ellipsoid can be created against the monitoring center. At last, we focus on protecting privacy in a network of agents who need to calculate a polynomial over their neighbors’ private values. Having this capability, the agents are enabled to calculate a wider range of functions beyond affine functions. We present an algorithm that accurately perform these calculations while keeping everyone’s data private. The method uses a new way to handle polynomials and includes cryptographic tools like Paillier encryption and secret sharing. The algorithm works across the entire network, uses minimal communication, and supports various types of functions.

Dissertation