Abstract
The blockchain technology has been extensively studied to enable distributed and tamper-proof data processing in federated learning (FL). Most existing blockchain assisted FL (BFL) frameworks have employed a third-party blockchain network to decentralize the model aggregation process. However, the decentralized model aggregation is vulnerable to pooling and collusion attacks from the third-party blockchain network. Driven by this issue, we propose a novel BFL framework that features the integration of training and mining at the client side. In our framework, the client first transmits its trained model to other clients, performs global aggregation upon receiving others' models, and competes to mine a block for aggregated model verification without the intervention of any third-party blockchain network. Considering the model transmission over time-varying wireless channels, we propose a dynamic training client scheduling to meet stringent latency requirement in FL, where clients with qualified channel conditions are scheduled to train their models in each communication round. Furthermore, we formulate a joint optimization problem of the training client scheduling and dynamic resource allocation (i.e., the transmit and computation power at the client side) under the constraint of long-term time-average (LTA) energy consumption. The objective of this optimization problem is to maximize the LTA training data size and thereby optimize the learning performance of FL. To this end, we obtain the closed-form solution by using the Lyapunov optimization method. Our experimental results show that the optimal solution can outperform baseline schemes in terms of learning accuracy and convergence time.