Optimal Privacy-aware Design of Networked Control Systems: An Information-theoretic Approach

Project: Research

View graph of relations

Description

Privacy is a major concern for the human users of networked control systems, e.g., smart buildings and intelligent transportation networks, since networked control systems increasingly rely on third-parties for performing computational tasks, e.g., cloud-computing units employed for monitoring, control, and fault detection purposes. These computational entities have legitimate access to the sensor measurements (or state estimates) of systems and may attempt to infer some (possibly private) information of the human users of these systems. For example, a cloud-based controller, with access to the temperature and CO2 measurements of a building, can potentially infer the occupancy information, i.e., the number of occupants in the building, which could be highly private. In many such cases especially in military and high-tech industry, privacy breaches have detrimental impacts on system designers, users as well as service providers. This proposal will investigate the optimal design of networked control systems under privacy constraints. First, we will introduce the notion of conditional entropy as a privacy measure for networked control systems. The conditional entropy determines how accurately private information can be inferred from the state estimates or sensor measurements. Using information-theoretic identities and inequalities, we will quantify the privacy level of a networked control system in various settings. We will also study the impacts of system parameters on the privacy level. These results can be used for privacy risk assessment, i.e., specifying the ability of a third-party in inferring private information of a system. We will next study the estimator and controller design problems for networked control systems under privacy-protection requirements. Here, we will formulate an entropypenalized approach for the optimal privacy-aware estimator and controller design problems. In this approach, the conditional entropy will be imposed as a penalty on the objective function of the estimator (controller) design problem. We cast the privacy-aware estimator (controller) design problem as a dynamic optimization problem, and investigate the structural properties of the optimal privacy-aware estimators and controllers. We will then utilize the structural results to develop computationally efficient algorithms for the privacy-aware design of networked control systems and to analyze the optimal trade-off between privacy and performance in these systems. We will finally integrate all the above-developed components to validate the entire framework toward practical engineering applications. To this end, we will design privacy-aware feedback control strategies for heating, ventilation, and air-conditioning (HVAC) systems using the entropy-penalized approach, and further examine the performances of these strategies using an HVAC simulator. 

Detail(s)

Project number9048210
Grant typeECS
StatusActive
Effective start/end date1/01/22 → …