RCEFL: Reputation-based Communication-aware Optimization of Federated Learning
Xun Liu, Jian Yang, Wenke Yuan, Yunpeng Hou, Feng Yang, Xiaofeng Jiang · 2024
Federated learning (FL) is a privacy-preserving distributed machine learning paradigm, where multiple devices holding local datasets collaborate to train a general global model without disclosing data. Instead of submitting raw data as in centralized training, devices send local gradients to the server for aggregation, which introduces the challenge of communication overhead. In each aggregation round, the server broadcasts the global model to all devices, while the devices train the model using local data and submit gradients to the server, which may lead to server congestion. Several works in communication optimization design device selection strategies to optimize model convergence efficiency, but they usually only consider the impact of the number of aggregation devices on the model convergence rate and communication cost, ignoring the impact of data heterogeneity on convergence rate. In this paper, we propose a reputation-based communication-aware federated learning optimization scheme named RCEFL. In RCEFL, we design a reputation evluation algorithm to predict the influence of device gradients on model convergence. Leveraging the reputation, we propose an efficient device selection and bandwidth allocation scheme to reduce model convergence time. Experiments demonstrate that our proposed RCEFL effectively reduces the convergence time of FL.