Speaker
Description
Hierarchical federated learning (HFL) can improve scalability and communication efficiency by combining client updates at intermediate aggregators before global coordination. However, grouping does not eliminate existing data and resource differences, and intermediate aggregation can reduce the client-level information available to the cloud.
Drawing on a challenge-oriented survey of 39 focal studies, this poster examines three connected post-grouping challenges: learning drift within and across groups, stale updates and uneven participation, and limited member-level global observability. We connect these challenges to their underlying causes and use representative methods to explain how existing approaches respond and which limitations remain. Particular attention is given to how group summaries may conceal participation, freshness, and update diversity, depending on the aggregation and reporting protocol.
The analysis highlights that effective responses depend on where information is available, which decisions each tier can control, and which parties are trusted. It exposes trade-offs among learning quality, communication cost, privacy, and robustness, and motivates coordination and adaptation strategies that account for information availability and practical grouping constraints.