Communication-efficient Federated Learning with an Event-triggering Strategy

With the development of artificial intelligence, data has become one of the most important resources. However, the emphasis on data privacy has brought great obstacles to the further development of machine learning algorithms. Federated learning keeps the local data set from being exposed to the outside world by passing model parameters during the training process, so communication cost has become a major bottleneck in the training process of federated learning. In this paper, to reduce useless communications during parameter upload, we propose a new federated learning algorithm with an event-triggering strategy based on model performance. Through this strategy, a certain number of communications have been reduced, so as to reduce the communication cost in training. In order to verify the performance of the proposed algorithm, experiments have been implemented on MNIST. The experimental results show that the proposed algorithm can reduce the communication cost by 38% at the same level of accuracy compared with the federated average. Compared with other event-triggering strategies, it can have less communications at the same level of accuracy.