Federated learning through model gossiping in wireless sensor networks

Federated learning (FL) has been recently proposed to achieve distributed learning in communication in an efficient and privacy-preserving way. Unfortunately FL cannot be applied as it is in wireless sensor networks, because it requires a large number of transmissions of the model parameters and because it involves huge energy and communication resource consumption at nodes that are physically closer to the point where models calculated by the federated learners are aggregated. In this paper, such claim will be demonstrated and we will investigate how gossiping, combined with FL, can be used to achieve higher energy efficiency and bandwidth saving. Early results show that the gossiping can help in improving significantly the performance in terms of resource efficiency.