TLDR: FedOC is a new framework for multi-server Federated Learning in wireless edge networks where edge server coverage areas overlap. It utilizes “overlapping clients” in two ways: as Relay Overlapping Clients (ROCs) to forward models between neighboring edge servers in real-time, and as Normal Overlapping Clients (NOCs) to dynamically select the earliest-received model for local training. This approach significantly reduces communication latency, addresses data heterogeneity, and achieves faster convergence and higher accuracy compared to existing methods, making it ideal for latency-sensitive edge applications.
Machine Learning (ML) has driven incredible advancements in various data-driven applications, from personalized medicine to autonomous driving. However, the traditional approach of centralizing data for model training often clashes with growing privacy concerns, especially when dealing with sensitive information like personal images or medical records. Federated Learning (FL) emerged as a powerful solution to this dilemma, allowing a shared model to be trained across data distributed on multiple client devices, thereby preserving user privacy while enabling scalable learning.
While single-server FL systems have been instrumental, they face significant communication bottlenecks and long-range latency, particularly in large-scale deployments. This has led to the development of multi-server FL architectures, which aim to alleviate these communication delays by distributing the workload across multiple edge servers (ESs).
A key observation in modern 5G and beyond wireless networks is that the coverage areas of different edge servers often overlap. This means that some clients, termed ‘overlapping clients’ (OCs), can access and communicate with multiple edge servers simultaneously. Most existing multi-server FL methods, however, typically assume that edge server coverage areas are distinct and do not account for these overlapping regions.
Introducing FedOC: A Novel Approach to Multi-Server FL
A new framework called FedOC (Federated learning with Overlapping Clients) has been proposed to fully harness the potential of these overlapping clients. FedOC introduces a novel architecture that allows overlapping clients to play dual, crucial roles in enhancing the efficiency and performance of federated learning in wireless edge networks. For more in-depth technical details, you can refer to the original research paper.
In FedOC, overlapping clients are categorized into two types:
- Relay Overlapping Clients (ROCs): These clients act as real-time intermediaries, forwarding aggregated edge models between neighboring edge servers. This mechanism facilitates direct model sharing and decentralized aggregation among different ESs, significantly speeding up the training process by allowing information to propagate across cells without constant reliance on a central cloud server.
- Normal Overlapping Clients (NOCs): These clients dynamically select their initial model for local training. Instead of waiting for all available edge models, NOCs choose the earliest-received model from any connected ES. This ‘Fastest Selection Strategy’ is crucial for meeting real-time requirements in latency-sensitive environments and helps mitigate data heterogeneity across different edge regions.
How FedOC Works
The overall workflow of FedOC proceeds in rounds. In each round, clients (both local and overlapping) train their local models based on the earliest received edge model. They then transmit their updated models to their respective edge servers for an initial aggregation. Following this, each edge server transmits its aggregated model to neighboring ESs via the designated ROCs. Upon receiving these relayed models, each ES performs a second aggregation, incorporating the models from its neighbors, and then broadcasts this newly updated model to its covered clients. This decentralized model dissemination indirectly achieves inter-cell model fusion and accelerates the training process.
Key Advantages and Performance
FedOC offers several significant advantages. It enables cross-server collaboration through OCs, facilitates real-time model relaying between ESs, and incorporates latency-aware training strategies for OCs. Theoretical analysis supports its convergence properties, and extensive experimental results demonstrate remarkable performance gains compared to existing methods.
Simulations conducted using datasets like MNIST and CIFAR-10 show that FedOC achieves faster convergence and higher training efficiency. For instance, on the harder CIFAR-10 task, FedOC-Fastest achieved a final accuracy approximately 9–12 percentage points higher than baseline methods. These gains are particularly pronounced in scenarios with minimal overlapping clients and even more so when the number of overlapping clients increases. The framework also proves robust and scalable when extended to networks with more edge servers, consistently delivering superior performance even under highly skewed data distributions.
Also Read:
- Optimizing Large Language Model Training in Mobile Edge Networks with CollaPipe
- Enhancing Trust and Performance in Healthcare Federated Learning
Conclusion
FedOC presents a compelling solution for federated learning in modern wireless edge networks. By intelligently leveraging overlapping clients as both communication relays and adaptive training participants, it effectively addresses critical challenges such as communication latency and data heterogeneity. This makes FedOC particularly well-suited for latency-sensitive applications in edge environments, where high client-server delays and non-IID (non-independently and identically distributed) data are common.
