Towards Private Data-driven Control
Andreea B. Alexandru, Anastasios Tsiamis, George J. Pappas
Abstract
Control as a Service (CaaS) is becoming a reality- particularly in the case of building automation and smart grid management. Often, the control algorithms in CaaS focus on controlling the client's system directly from input-output data, since the system's model might be private or unavailable. Therefore, large quantities of data collected from the client need to be uploaded to a cloud server. This data can be used by a malevolent cloud service provider to infer sensitive information about the client and mount attacks. In this paper, we co-design a solution that interlaces control and privacy. Our goal is to perform online data-driven control on encrypted input-output data, while maintaining the privacy of the client's uploaded data, desired setpoint and control actions. We design our control algorithm based on results from the behavioral framework, which is more encryption-friendly compared to other classical frameworks. We obtain privacy by using a leveled homomorphic encryption scheme to enable the cloud to perform complex computations on the client's encrypted data. Finally, we achieve efficiency by manipulating the tasks required by the control algorithm such that they only involve arithmetic circuits, as well as by leveraging parallelization and ciphertext packing.