MGM-4-FL: Combining federated learning and model gossiping in WSNs

Application of Federated learning (FL) as it is, in Wireless Sensor Networks (WSNs) in impractical. In fact, it requires the transmission of a large number of model parameters. Also, it implies the consumption of a massive amount of energy and communication resources at nodes located closer to the network elements that aggregate the models calculated by the federated learners. This is a well known problem in WSN called funneling effect. Solutions have been recently proposed that decentralize FL operations further by exploiting multihop communications. Nevertheless, the proposed approaches do not focus on the networking operations needed to support such decentralization in an efficient manner. Objective of this paper is to fill this gap by proposing an integrated networking/learning scheme, named Model Gossiping Method for FL, that supports distributed FL efficiently in wireless sensor networks.