Training Data Cost Ratio Optimization for Federated Learning in Cellular Internet of Things

The cellular Internet of Things (IoT) enhanced by federated learning (FL) is a potential paradigm to leverage the vast amount of data generated by the IoT devices and offer various intelligent applications. Through its distributed learning manner, the privacy and delay problems of the learning process are well handled. Nevertheless, in the cellular IoT, FL requires multiple rounds of model parameters exchanging between the parameter server and multiple clients over unstable wireless links, which largely constrains the communication efficiency. Regarding this problem, we propose the training data cost ratio to evaluate the communication efficiency, and then by maximizing this metric, client scheduling, transmitting power, and bandwidth are jointly formulated. The formulated problem is decomposed via problem transformation and derivations, and then, the Lagrange method and greedy-based algorithms are developed to solve the subproblems efficiently. Simulation results verify the advantages of our algorithm in communication efficiency improvement. Moreover, it reveals that the proposed metric and joint optimization substantially obtain superior tradeoff between learning performance and resource consumption compared to the client number oriented optimization.