Incentivizing task offloading in IoT: A distributed auctions-based DRL approach

Abstract

Federated Learning (FL) has emerged as a powerful tool for leveraging the vast quantities of data generated by Internet of Things (IoT) devices. Its advantage lies in its ability to preserve participants’ data privacy through keeping it local. In addition, FL alleviates the communication overhead in cloud-centric ML for IoT approaches, through sharing model updates instead of large raw data, which optimizes bandwidth use. Furthermore, decentralized FL has been useful in addressing security concerns typical in conventional settings. Nevertheless, in such zero-trust scenarios where there is no central coordinator, nodes exhibit reluctance to participate due to the lack of clear rewards and trust issues. Additionally, constrained FL clients may abandon their tasks, which negatively impacts learning performance. In this paper, we propose a double-incentive FL approach to address the dual challenge of node reluctance and task offloading in a fully distributed FL-based IoT network. We introduce an auction-based offloading scheme to handle task abandonment. Multi-Agent Deep Reinforcement Learning (MADRL) is leveraged to build a bidding strategy with long-term optimization of system and individual utilities. We also present a client filtering and rewarding algorithm based on a reputation model. Our objective is to promote truthfulness and enhance resilience against malicious nodes, while improving energy efficiency and accuracy. By employing a REINFORCE-based scheme for offloading, our approach demonstrates a superior trade-off between energy efficiency and accuracy, as well as resilience to malicious behavior. Furthermore, empirical results highlight its performance in terms of truthfulness, despite the uncertain and opaque nature of the environment.