Grouped federated learning for time-sensitive tasks in industrial IoTs

The Industrial Internet of Things (IoT) is a crucial part of Industry 4.0 and constantly calls for intelligence to improve productivity. Meanwhile, the federated learning scheme has recently been proposed to provide both distributed learning ability and inherent privacy protection functions. Despite its original advantages, the unique features of industrial IoT posed several challenges to federated learning regarding efficiency and scalability. Due to the heterogeneity of data under IoT and the large difference in computing efficiency of each device, the training time of federated learning will be affected by stragglers. To this end, an effective device selection mechanism is urgently needed to improve the training efficiency of global model. Additionally, the communication coordination problem among various devices should be considered to accelerate global convergence. To solve the abovementioned problems, we proposed an Upper Confidence Bound (UCB)-based grouping federated learning (UGFL) where a data scheduling method effectively reduces the stragglers. We combined Lyapunov optimization with the UCB (Garivier and Moulines in International Conference on Algorithmic Learning Theory, pp. 178–188, 2011) algorithm to design an efficient device selection method to choose the best device to participate in each global epoch. In addition, we theoretically analyzed the upper bound of convergence of our proposed method. Experimental results show that our method has higher accuracy and faster convergence speed than the existing methods.