Standard Federated Learning (FL) is built on a "star" topology: a single central server coordinates thousands or millions of edge devices. While conceptually simple, this architecture creates a massive communication and computation bottleneck. As the number of participants scales, the central server becomes a single point of congestion, struggling with simultaneous model uploads, high latency, and the sheer processing power required to aggregate high-dimensional updates.
To move beyond these limitations, researchers are reimagining the network structure of FL, shifting from centralized hubs to hierarchical trees and decentralized meshes.
The Failure of the Star Topology at Scale
In a traditional star network, every device must maintain a long-range uplink to a distant cloud server. As noted by Mehdi Salehi Heydar Abad and colleagues in Hierarchical Federated Learning ACROSS Heterogeneous Cellular Networks, this results in severe network congestion and energy drain.
The bottleneck isn't just about bandwidth; it's also about the server's computational limits. shay vargaftik and the team behind EDEN point out that if a server must perform complex inverse transformations or decompressions for every individual client, the processing time scales linearly with the participant pool. In a world of millions of devices, the central aggregator simply cannot keep up.
Hierarchical Aggregation: Divide and Conquer
One dominant solution is Hierarchical Federated Learning (HFL). Instead of every phone talking to the cloud, devices are organized into clusters managed by intermediate "edge" or "fog" nodes—such as small cell base stations or local routers.
Seyyedali Hosseinalipour et al. demonstrate in Multi-Stage Hybrid Federated Learning over Large-Scale Wireless Fog Networks that by leveraging these intermediate layers, systems can reduce network traffic by up to 80%. The intuition is simple: aggregate locally and frequently, then sync globally and sparingly. This "multi-tier" approach is further refined by Sen Fu in Hierarchical Federated Learning with Momentum Acceleration in Multi-Tier Networks, which uses momentum at every level of the hierarchy to speed up convergence, slashing training time by 70%.
Decentralized and Ring-Based Architectures
If the central server is the problem, some researchers argue we should remove it entirely. Fully decentralized FL relies on peer-to-peer (P2P) communication, where nodes share updates with their immediate neighbors.
- Ring Topologies: Jin‐Woo Lee and the authors of TornadoAggregate propose a ring-based architecture. Instead of a star, clients pass updates in a circle. This provides near-linear communication scalability, though it requires careful management to prevent a single malicious "Byzantine" node from poisoning the entire chain, a challenge addressed by Minghong Fang in Byzantine-Robust Federated Learning over Ring-All-Reduce Distributed Computing.
- P2P Graphs: Ye Yuan introduces DeFed, a framework that mathematically proves P2P systems can match the convergence rates of centralized ones while being far more resilient to network failures. Similarly, Felix Mulitze presents MAR-FL, which uses iterative group-based aggregation to reduce communication overhead from $O(N^2)$ to $O(N \log N)$.
Serverless and Multi-Server Coordination
The transition toward "serverless" FL aims to eliminate the operational complexity of maintaining a dedicated aggregation hub. Xidong Wu and Zhengmian Hu developed SLATE, a serverless algorithm that treats model aggregation as a decentralized stochastic optimization problem. In this model, the "server" is replaced by shared cloud storage or client-side logic, as seen in the flwr-serverless framework by Sanjeev V. Namjoshi.
In wireless environments, this decentralization can even turn clients into infrastructure. Yun Ji and colleagues introduced FedOC, which uses clients located in overlapping server coverage areas as "smart relays." These relays forward models between multiple servers, effectively weaving a mesh that bypasses the congestion of any single central point.
Why It Works: Shifting the Burden
The core intuition behind this research direction is the spatial distribution of overhead. By moving aggregation closer to the data source (Hierarchical) or spreading it across the entire network (Decentralized), the system:
- Reduces Latency: Communication happens over shorter, faster local links.
- Increases Bandwidth: The total available bandwidth of the network is used, rather than the narrow pipe leading to a single server.
- Enhances Robustness: The system no longer has a single point of failure.
As federated systems move toward millions of participants, the "star" is fading, replaced by more organic, tiered, and peer-driven architectures that mirror the decentralized nature of the data they seek to learn from.