Communication-Efficient Federated Learning under Dynamic Device Arrival and Departure: Convergence Analysis and Algorithm Design

Researchers have made advancements in federated learning (FL) by addressing challenges posed by dynamic device arrival and departure, achieving convergence speedups of an order of magnitude or more compared to baselines^[1].

Federated learning enables multiple clients to collaboratively train a shared-task model while maintaining privacy^[2]. However, traditional FL approaches assume a fixed device set, which is not always the case in real-world scenarios where devices may join or leave the network. This dynamic setting introduces challenges such as evolving optimization objectives and reduced global model effectiveness. To address these challenges, a new approach provides a convergence analysis for FL under a dynamic device set, taking into account factors like gradient noise, local training iterations, and data heterogeneity. The proposed model initialization algorithm computes a weighted average of previous global models based on gradient similarity, prioritizing models trained on data distributions that closely align with the current device set. Experiments have shown that this approach achieves significant convergence speedups compared to existing methods^[1]. Furthermore, a data-centric review of FL highlights the importance of addressing data-related vulnerabilities and challenges, which can affect convergence. The review analyzes non-IID data into measurable traits and ranks their influence on convergence, emphasizing the need for careful consideration of data-related issues in FL^[2].