On Dynamic Resource Allocation for Blockchain Assisted Federated Learning over Wireless Channels

Blockchain assisted federated learning (BFL) has been widely studied as a promising technology to process data at the network edge in a distributed manner. However, the study of BFL involves key challenges, including resource allocation and client scheduling. In this paper, we propose a BFL framework consisting of multiple clients, where the roles of clients include local model training, wireless uploading, and block mining in each round. First, we develop a renewal BFL framework to study the long-term system performance under time-varying fading channels. Second, in order to speed up the BFL process with limited communication, computation and energy resources, we propose a dynamic resource allocation and client scheduling (DRACS) algorithm based on Lyapunov optimization to maximize the training data size under energy consumption constraints, by jointly optimizing the allocation of communication, computation, and energy resources. For the DRACS algorithm, we characterize a trade-off of [$\mathcal{O}(1/V)$, $\mathcal{O}(\sqrt{V})$] between the training data size and energy consumption to balance the maximization of training data size and the minimization of energy consumption. Our experimental results show that the DRACS algorithm can provide both higher learning accuracy and faster convergence with limited time and energy based on the MNIST and Fashion-MNIST datasets.