ABSTRACT
Federated learning (FL) is a key technique for training machine learning models in a collaborative and privacy-promoting fashion. FL is based on the iterative training of a shared model on the local data of heterogeneous participants, and only requires the upload of resulting updates to a centralized entity for aggregation. Nevertheless, the extensive computation and communication costs linked to scaling this technique raise several concerns and challenges. In this chapter, we discuss different considerations to achieve green FL alongside key enablers and open challenges. We begin by introducing a general template of FL and its key challenges. Then, we discuss techniques for initializing lightweight models. Next, we discuss the main steps in the iterative process, namely client selection, local training, model upload and aggregation, and personalization. We review key considerations and recent research efforts toward green FL for each step of the process. Finally, we discuss future research directions and open challenges.
