Joint Provisioning of QoS and Privacy with Federated Learning

Abstract

Combined optimization of Quality of Service (QoS) and security/privacy has always been an overlooked topic among IT professionals. The main reason for this is the difficulty to formulate optimization problems with the various conflicting QoS, security, and privacy attributes. There is no closed-form solution to such an optimization problem and this is why there was a lack of a seamless integration of QoS and privacy in the literature for many decades. In an emerging smart society, data privacy has become a top priority, particularly in the healthcare domain, which deals with massive patient records, health data, and imaging data from a myriad of modalities. Storing and sharing such data across hospital networks requires not only privacy but also presents a considerable challenge in terms of QoS overheads due to the massive bandwidth needed. The closed-source medical record storage and sharing platforms contribute to this challenge even more. We use this as a motivational use case in this chapter to demonstrate the need for seamless QoS and privacy provisioning of medical data exchange among various stakeholders. In this vein, we discuss the applicability of a distributed, decentralized learning framework called federated learning that provides a new paradigm for medical record platforms, biomedical equipment, and even resource-constrained IoT (Internet of Things) devices to participate in collaborative AI (artificial intelligence) model building. The federated learning framework provides two advantages, one from the QoS point of view and the other from the privacy-preserving aspect. As for the QoS assurance, federated learning techniques typically share the wisdom in terms of the locally developed AI model parameters instead of the raw big data, which directly improves the communication delay and bandwidth overheads. On the other hand, it preserves the users’ data privacy by eliminating the need to send the original data (which could be sensitive) with remote cloud or third-party stakeholders. In this chapter, we describe an asynchronously weight updating federated learning technique which can further improve the performance of federated learning setups in terms of both QoS and privacy preservation at the same time.