Distributed Federated Deep Learning in Clustered Internet of Things Wireless Networks With Data Similarity-Based Client Participation

Evangelia Fragkou, Eleftheria Chini, Maria Papadopoulou, Dimitrios Papakostas, Dimitrios Katsaros, Schahram Dustdar

Abstract: Federated deep learning is the method of choice for performing deep learning in environments where data sharing is not allowed due to privacy/security issues. However, all of the solutions based exclusively or substantially on the existence of a coordinating server are not a fit for wireless Internet of Things environments operating as ad hoc networks due to the excessive communication overhead. This article develops the DISCREETER distributed protocol to perform distributed federated learning, fully addressing the requirements and peculiarities of the wireless ad hoc IoT environment. The algorithm is based on a hierarchical organization of the underlying network along with a generic principled way to select the clients that will participate in the federation based on the similarity of the data they produce. We evaluate the proposed methodology on real data, targeting a time series prediction task, using a recurrent neural network in each node as the learning entity. The obtained results attest to the design merits of our approach.