Abstract: Two-step aggregation is introduced to facilitate scalable federated learning (SFL) over passive optical networks (PONs). Results reveal that the SFL keeps the required PON upstream bandwidth constant regardless of the number of involved clients, while bringing ~10% learning accuracy improvement.
Introduction
Federated Learning (FL) has emerged as a critical paradigm for training artificial intelligence models while preserving data privacy. By allowing clients to train models locally and only share parameter updates, FL avoids the need to centralize sensitive raw data. However, deploying FL over Passive Optical Networks (PONs)—the backbone of modern fiber-to-the-home infrastructure—presents significant challenges. PONs typically offer asymmetric bandwidth, with upstream capacity being much more constrained than downstream. In a standard synchronous FL setup, hundreds of clients attempting to upload multi-megabyte model updates simultaneously lead to severe network congestion. This congestion results in "stragglers"—clients whose updates fail to reach the server within a required time window—thereby degrading the final model accuracy.
Proposed Method: Scalable Federated Learning (SFL)
To address these scalability issues, this work proposes a Scalable Federated Learning (SFL) architecture utilizing a "two-step aggregation" method. This approach shifts the initial aggregation process from the central office to the network edge.
In the standard approach, every client transmits its full model update through an Optical Network Unit (ONU) and across the PON link to a Central Parameter Server (CPS). In contrast, the SFL approach introduces an Aggregation Function (AF) at the ONU level:
- Edge Aggregation: Each ONU collects model updates from its locally connected clients (e.g., via Wi-Fi or base stations) and performs a weighted addition of these parameters.
- Final Aggregation: The ONU then transmits a single, aggregated update to the CPS.
By aggregating at the ONU, the traffic on the primary PON link becomes invariant to the number of clients; it depends solely on the number of ONUs in the network.
Experimental Evaluation
The SFL framework was evaluated using the FEMNIST dataset, a standard benchmark for federated learning consisting of handwritten characters partitioned by writer. The model used was a Convolutional Neural Network (CNN) with two 5x5 convolution layers, resulting in a model update size of approximately 26.416 Mbits per client. The simulation environment modeled 16 ONUs and varying numbers of clients (N=48 and N=128).
Key Results and Discussion
The experimental results demonstrate that SFL effectively decouples network performance from the number of participating clients.
- Bandwidth Efficiency: As shown in the results, SFL maintains a constant upstream bandwidth requirement. Compared to the benchmark, SFL reduced upstream bandwidth usage by 66.7% for 48 clients and 87.5% for 128 clients.
- Participation Stability: In the benchmark scenario, network congestion caused the number of successfully involved clients to fluctuate wildly, often dropping to near zero as the total client count increased. SFL maintained high and stable participation rates because the bottleneck at the PON link was eliminated.
- Accuracy Improvement: The increased and stable participation directly translated to better model performance. The learning accuracy for the FEMNIST task improved from 0.77 in the benchmark to 0.85 with SFL—a relative increase of approximately 10%.
Conclusion and Future Work
Scalable Federated Learning via two-step aggregation successfully mitigates the upstream bandwidth bottleneck in PONs. By performing initial aggregation at the ONU, the system ensures that adding more clients improves model accuracy without compromising network stability. Future research will focus on the hardware implementation of Aggregation Functions within "Smart ONUs" and the integration of these processes into existing Layer 2 MAC protocols.
Summary updated