Communication-Efficient Hierarchical Federated Learning for IoT Heterogeneous Systems with Imbalanced Data

• Privacy-sensitive medical and IoT data make forwarding data to a centralized entity impractical, while centralized processing burdens wireless networks.Source: §1 · p. 1
• Repeated full-model updates to a central server increase communication overhead and convergence time, especially with non-IID data.Source: §1 · p. 2
• The proposed hierarchy assigns users to edge nodes using location, data distributions, and communication constraints, with edge aggregation and periodic central synchronization.Source: §1 · p. 2
• The work formulates user assignment and resource allocation around data distributions, latency, bandwidth, and energy constraints, and proposes an algorithm considering wireless conditions.Source: §1 · p. 2

2.1 Single-layer Federated Learning

• Single-layer FL research spans convergence and resource optimization, user selection, and communication-efficient sparsification or compression.Source: §2.1 · p. 3
• Non-IID research includes sharing data to reduce skew, adapting aggregation frequency, selecting users, and compressing updates, but some approaches leave wireless allocation or hierarchical FL unaddressed.Source: §2.1 · p. 3

2.2 Hierarchical Federated Learning

• Hierarchical FL studies address convergence, user selection, and data imbalance, but prior work identified by the authors leaves practical wireless resource allocation constraints unaddressed.Source: §2.2 · p. 3
• The authors present their work as the first hierarchical-FL EU assignment approach to jointly consider user data distributions, communication characteristics, and resource allocation, with evaluation on two real-world health datasets.Source: §2.2 · pp. 3–4

3.1 Challenges of FL in healthcare systems

• Healthcare FL must contend with heterogeneous non-IID patient data, large participant populations, and limited bandwidth, computation, and energy resources.Source: §3.1 · p. 4

3.2 I-Care system architecture

• I-Care places patient-associated IoT devices and local learning at the EU layer, with edge nodes between users and the centralized server to reduce full-update traffic.Source: §3.2 · p. 4
• Within each EU group, FedSGD aggregates local models at the serving edge node; edge models then synchronize at the central server using FedAvg, which returns the global model to participants.Source: §3.2 · p. 4

4.1 Hierarchical FL problem formulation

• The learning objective is to minimize cross-entropy loss over the virtual global dataset, while EUs train locally and periodically synchronize weighted model averages.FedSGD is used for local updates; the virtual datasets clarify analysis and are not physically stored at edge nodes or the centralized server.Source: §4.1 · p. 5
• Hierarchical FL aggregates EU weights at edge nodes every T′ local-gradient steps, then aggregates edge models centrally every T′ × T steps.The centralized model is returned to edge nodes and EUs; synchronous operation waits for all EUs to finish local updates before communication.Source: §4.1 · pp. 5–6

4.2 Computation and Communication Latency Analysis

• The latency model includes EU-to-edge upload and download delays and local computation time, with computation depending on model size, dataset size, CPU cycles, and frequency.The local-iteration count is upper-bounded by O(log(1/ϵ)); the modeled delay focuses on EU-edge links because edge-to-server delay is treated as constant across associations.Source: §4.2 · p. 6

4.3 Transmission Energy Consumption Analysis

• Transmission energy is modeled from the EU-edge wireless channel, transmission rate, and allocated bandwidth for sending local updates.The model assumes a deterministic path-loss channel and excludes processing energy because it is described as smaller and unaffected by assignment or bandwidth allocation.Source: §4.3 · pp. 6–7

CATION

• The problem targets resource-constrained systems with globally imbalanced data and asks how to assign EUs to edges and use available resources while training adequately.Resources include computation and communication time, energy, and wireless bandwidth.Source: CATION · p. 7

5.1 Problem Formulation

• The optimization minimizes KLD between edge-node class distributions and a uniform reference while accounting for delay, transmission-energy, bandwidth, and connectivity constraints.The assignment variables determine EU-edge associations; bandwidth variables allocate capacity to assigned EUs.Source: §5.1 · pp. 7–8
• The formulation is an NP-complete integer program, motivating a solution that reformulates its objective and relaxes it to a linear program.The relaxation removes the bandwidth constraint before rounding and subsequent bandwidth allocation.Source: §5.1 · p. 8, §5.2 · pp. 8–9

5.2 Proposed Solution

• The EARA algorithm first solves a relaxed linear program and rounds assignments to satisfy integer constraints, then allocates bandwidth while preserving the KLD minimization.Bandwidth is assigned by EU importance to KLD reduction, providing the minimum needed to meet latency constraints until the edge bandwidth budget is exhausted.Source: §5.2 · pp. 8–9
• EARA supports single connectivity to one edge or dual connectivity to one or two edges, and ranks assigned EUs by their contribution to KLD minimization.Dual connectivity relaxes the one-edge constraint; the bandwidth allocation procedure prioritizes EUs with data classes different from those available at an edge.Source: §5.2 · p. 9

6.1 Environment setup

• Evaluation uses Heartbeat ECG and Seizure EEG datasets with randomly distributed non-IID EU data; EARA-SCA/DCA are compared with nearest-edge DBA and centralized learning.Source: §6 · p. 9, §6.1 · pp. 9–10, §6.2 · p. 10
• EARA centrally computes EU assignments, then edge nodes sort assigned users by weight and allocate the minimum bandwidth satisfying the constraint.Source: §6.1 · p. 10

6.2 Simulation results

• Under DBA, accuracy is below centralized learning even with all EU updates; dropping users further hurts, especially when one or two classes disappear.Source: §6.2 · pp. 10–11, Figure 3 · p. 11
• EARA reduces edge-level KLD versus DBA, with EARA-DCA lowest; at large distances, energy constraints push assignments toward nearer edges, making EARA converge toward DBA.Source: Figure 4 · p. 11, §6.2 · p. 11
• 75–85% fewer communication rounds than DBA at the same model accuracy; EARA-DCA also outperforms centralized learning in accuracy.Figure 5 compares classification accuracy across communication rounds on both datasets.Source: §6.2 · pp. 11–12, Figure 5 · p. 12
• At 90% Heartbeat accuracy, EARA-SCA cuts per-EU traffic 50% versus DBA; EARA-DCA cuts SC-EU traffic 73% and DC-EU traffic 47% versus DBA.For DC-EUs, EARA-DCA traffic is 3% higher than EARA-SCA.Source: Figure 6 · p. 12, §6.2 · p. 12

7 CONCLUSION

• The scheme brings edge-level user data distributions closer to uniform, alleviating local imbalanced and non-IID data effects and reducing KLD versus distance-based assignment.The paper links this reduction in KLD to fewer global aggregation rounds.Source: §7 · pp. 12–13
• 75–85% fewer communication rounds between edge nodes and the centralized server, compared with state-of-the-art solutions, while preserving the same model accuracy.The reduction follows from decreasing global aggregation rounds on the centralized server.Source: §7 · p. 13
• Tight synchronization among nodes and the presence of a centralized server remain challenges for making FL suitable for highly dynamic scenarios such as vehicular applications.Future work could develop online algorithms to select outstanding users while minimizing communication overhead.Source: §7 · p. 13