Cooperation & competition : mechanism design for federated optimization of edge intelligence

Abstract

With the proliferation of Internet of Things (IoT) applications, a huge amount of data is generated at the network edge. Due to bandwidth, storage, and most importantly privacy concerns, it is impractical to move the local data to the cloud for centralized analytics model training. Federated learning (FL) has been widely recognized as a promising approach that enables individual edge devices (also known as “clients”) to train a global model cooperatively without exposing their data and other private information. However, although FL has substantial advantages, it still faces the following chal­lenges: First, due to the significant energy consumption for joining FL, the non­-independent and identically distributed (Non-IID) data and heterogeneous hardware resources, the clients may be reluctant to participate in the FL without proper re­wards from well-designed incentive mechanisms. Second, the Non-IID data distribu­tion among extensive clients makes it impossible for a single global model to adapt to the requirements of all clients simultaneously. This naturally derives the demand of personalization on different clients, i.e., each client requests a personalized model that can perfectly fit their local Non-IID data. Third, the extensive clients involved in the FL system have strong dynamics. i.e., the new clients join and bring new knowl­edge, while the old ones exit and leave obsolete knowledge in the system. Normally, a fixed-size model has an upper limit on its knowledge capacity, i.e., it cannot learn new knowledge indefinitely. Therefore, under the cooperation of all clients, model un­learning for obsolete knowledge is highly required in the FL system to accommodate its strong dynamics. In this thesis, we investigated the problems and challenges of these issues from the perspectives of game theory and mechanism design, where the corresponding solutions are proposed to handle them. For the first issue, we present long-term online VCG auction incentive mechanisms based on deep reinforcement learning, which can adap­tively assign proper rewards to clients with different resource and data conditions. It considers several vital economic properties to guarantee a sustainable environment for the long-term development of the FL system. For the second issue, we propose a multiwise collaboration framework based on cooperative game theory, which only en­courages clients with relevant data distribution for collaboration and trains their own personalized model. For the third issue, we take the early step to comprehensively investigate the machine unlearning paradigm in the context of FL (i.e., federated unlearning) and thereby propose a general pipeline for federated unlearning based on stochastic gradient ascent (SGA) and client cooperation. We conduct extensive experiments to show the remarkable performance improvement of our proposed methods compared with the existing methods on various datasets and settings.