Hierarchical Federated Learning: Approaches, Applications, and Open Challenges

Federated learning (FL) is a machine learning (ML) paradigm that enables models to be trained collaboratively across distributed clients without breaking privacy. By allowing data to remain on clients’ devices and only sharing model updates, FL can preserve privacy while unlocking the potential for collective intelligence for a wide range of applications. Nevertheless, FL encounters several challenges. First, its two-tier infrastructure (clients and a server) is incompatible with the most recent three-tier end-edge-cloud infrastructure, resulting in high latency and sometimes a lack of adaptation. In addition, FL also suffers from communication overhead due to the frequent updates required for long-range client-to-cloud server communication. Recent research has also shown that FL is prone to data leakage, as the model updates could be used to reconstruct sensitive data. Hierarchical federated learning (HFL) has emerged as a promising FL extension that introduced multi-level organizational structures to reduce the effect of these challenges. This paper provides a comprehensive overview of HFL, encompassing its methodologies, applications, and challenges. The paper also discusses open research directions and the future trajectory of HFL, emphasizing the need for cross-disciplinary collaboration.