Key Generation and Secure Coding in Communications and Private Learning

Abstract

The increasingly distributed nature of many current and future technologies has introduced many challenges for devices designed for such settings. Devices operating in such environments, such as Internet-of-Things (IoT), medical devices, connected vehicles, etc., typically have limited computational power and rely on batteries to operate. Therefore, efficiency is a paramount requirement for any algorithm designed to be implemented on these devices. Furthermore, these devices typically generate and collect huge amounts of extremely sensitive and personal data, such as health-related data, behavior-related data, etc. As a result, there is a need for security and privacy protections to guard against various attacks. Additionally, since these devices are typically resource-constrained, any algorithm or protocol needs to be efficient to enable its implementation on such devices. Efficient security and privacy solutions are essential to cope with, as well as enable, high deployment rate of such devices for various sensitive applications.

In this dissertation, efficient solutions for protecting the security and privacy of data generated by such devices are explored. Low-complexity protocols for generating secret keys in static environments, along with a formulation of threshold-secure coding with a shared key and corresponding coding schemes are presented. Additionally, algorithms for coded machine unlearning for regression problems are presented, as well as a new setup and algorithm for federated learning with opt-out differential privacy are presented and evaluated.