Optimal Federated-Learning-Based Resource Allocation Scheme for NOMA-Enabled Industrial Internet of Things

As an emerging collaborative approach, Federated Learning (FL) addresses the privacy leakage issue in traditional centralized machine learning (ML), which requires devices to upload raw data for model training. However, the limited communication and computing resources and more stringent real-time requirements have brought great challenges to the selection of equipment and resource allocation involved in training in the Industrial Internet of Things (IIoT). To tackle those challenges with large-scale distributed devices, this paper proposes an optimal federated learning-based resource allocation scheme for NOMA-enabled IIoT. In order to cope with the long communication distance and high communication overhead between large-scale distributed IIoT devices and cloud server, a hierarchical cooperative FL architecture based on NOMA is designed. To maximize the system throughput, the power allocation strategies of the inter sub-channel and intra sub-channel are designed under the constraints of maximizing transmit power and guaranteeing the minimum data rate. To achieve the tradeoff between the federated learning performance and the training cost, the device selection and bandwidth resource allocation are established as a joint optimization problem. Subsequently, a DRL based joint optimization algorithm is proposed to improve the federated learning utility value of the system. The experimental results demonstrate that our proposed scheme can facilitate efficient FL and obtain better performance than the benchmark methods.