Personalizing Federated Learning for Hierarchical Edge Networks with Non-IID Data

TL;DR

This work tackles hierarchical non-IIDness in 3-level Federated Learning by recognizing that edges can have distinct optimization goals due to location-based data distributions. It introduces Personal-ized Hierarchical Edge-enabled Federated Learning (PHE-FL), which creates edge-specific cloud-aggregated models and blends them with edge-aggregated models using a dynamic weighting parameter $\alpha$ to form personalized edge models at every round. Empirical evaluation across MNIST, Fashion-MNIST, and CIFAR-10 under four edge-level non-IID scenarios shows that PHE-FL achieves up to 83% higher accuracy than baselines and exhibits improved stability compared to FedAvg-based edge-cloud architectures. The approach maintains high performance across varying data complexity and hyperparameters, demonstrating robustness and practical potential for real-world edge networks with heterogeneous data distributions.

Abstract

Accommodating edge networks between IoT devices and the cloud server in Hierarchical Federated Learning (HFL) enhances communication efficiency without compromising data privacy. However, devices connected to the same edge often share geographic or contextual similarities, leading to varying edge-level data heterogeneity with different subsets of labels per edge, on top of device-level heterogeneity. This hierarchical non-Independent and Identically Distributed (non-IID) nature, which implies that each edge has its own optimization goal, has been overlooked in HFL research. Therefore, existing edge-accommodated HFL demonstrates inconsistent performance across edges in various hierarchical non-IID scenarios. To ensure robust performance with diverse edge-level non-IID data, we propose a Personalized Hierarchical Edge-enabled Federated Learning (PHE-FL), which personalizes each edge model to perform well on the unique class distributions specific to each edge. We evaluated PHE-FL across 4 scenarios with varying levels of edge-level non-IIDness, with extreme IoT device level non-IIDness. To accurately assess the effectiveness of our personalization approach, we deployed test sets on each edge server instead of the cloud server, and used both balanced and imbalanced test sets. Extensive experiments show that PHE-FL achieves up to 83 percent higher accuracy compared to existing federated learning approaches that incorporate edge networks, given the same number of training rounds. Moreover, PHE-FL exhibits improved stability, as evidenced by reduced accuracy fluctuations relative to the state-of-the-art FedAvg with two-level (edge and cloud) aggregation.

Figures (4)

• Figure 1: Personalized Hierarchical Edge-enabled Federated Learning (PHE-FL). After cloud aggregation, PHE-FL creates personalized models at the edge, allowing connected clients to receive distinct, edge-personalized models.
• Figure 2: OnlyEdge Workflow: (1) Cloud server initializes the global model for the first round. (2) Global model is distributed to the edge servers. (3) Each edge server either distributes the initial global model in the first round or the aggregated edge model in subsequent rounds. (4) IoT devices train the edge model with local data. (5) IoT devices send the model weights back to the edge server. (6) Edge servers aggregate the models from IoT devices. Repeat (3) to (6).
• Figure 3: Performance of EdgeCloud, OnlyEdge, and PHE-FL on datasets (MNIST, F-MNIST, CIFAR10) across distributions D1–D6. EdgeCloud had the lowest accuracy on D1 and D2, OnlyEdge on D5 and D6, while PHE-FL never recorded the lowest accuracy.
• Figure 4: Results of additional experiments testing PHE-FL's robustness with different hyperparameters, all conducted under the MNIST D2 imbalanced test scenario.