A distributed D2D clustering algorithm tailored for hierarchical federated learning in a multichannel ALOHA network

Abstract

In an age characterized by seamless interconnectivity, the quantity of Internet of Things (IoT) devices has experienced substantial growth in recent years, with projections indicating further expansion. In this context, Federated Learning (FL) plays an important role in the future of wireless communications, offering numerous advantages over traditional centralized learning approaches, including data privacy preservation, reduced bandwidth usage, improved accuracy, and customization. FL is being very used on IoT applications, where multiple devices will contribute with their learning over its local private data in order to build a global model, which will be further aggregated and trained by the Base Station (BS) só that it is redistributed back to the devices. The aim of this work is creating a Device-to-Device (D2D) clustering algorithm, named D2D Short Range Clustering (D2D-SRC), capable of optimizing a multichannel ALOHA communication within a hierarchical FL system, with this learning being the main motivation. Selecting an appropriate wireless protocol and data transmission method is crucial for FL, therefore it was adopted the multichannel ALOHA protocol due to its asynchronous nature and simple implementation compared to other protocols. Also, ALOHA is an interesting practical alternative to the over-the-air computation (AirComp), which is a technique much used in the literature in wireless communication systems to compute functions of distributed data directly through the air without decoding individual messages. The numerical results show that the novel D2D clustering approach proposed is capable of hitting high clustering rates, improving the general efficiency of the system by drastically reducing the achievable error, while allowing a BS to serve more devices using the same resources.