Horizontal Federated Learning

The fundamental prerequisite for HFL is that all participating clients must agree on a common data schema. Every base station or client node records the same set of features—such as Reference Signal Received Power (RSRP) , Signal-to-Interference-plus-Noise Ratio (SINR) , and throughput metrics—but for entirely different user equipment (UE) populations.

• Structural Homogeneity: The column names, data types, and semantic meanings are identical across all silos.
• Sample Heterogeneity: The row entries represent distinct, non-overlapping user groups.
• This alignment allows a global neural network architecture to be defined centrally and deployed to all clients without modification.

Data is partitioned by sample ID rather than by attribute. In a telecom context, Operator A and Operator B both record the same call detail records and channel quality indicators, but Operator A's data represents subscribers in Berlin while Operator B's data represents subscribers in Munich.

• No Entity Overlap: A specific IMSI (International Mobile Subscriber Identity) appears in only one client's dataset.
• Vertical Complement: This contrasts directly with Vertical Federated Learning, where the same users are described by different attributes across organizations.
• This topology naturally fits geographically distributed RAN deployments where user mobility is tracked within a single operator's network.

HFL is the native topology for the standard Federated Averaging algorithm. Because local models share an identical architecture, the central server can perform a weighted element-wise average of the model weights received from participating clients.

• Weighted Aggregation: Updates are typically weighted by the number of local training samples to account for dataset size imbalances.
• Convergence Stability: Shared feature distributions generally lead to smoother global loss landscapes compared to non-IID label distributions.
• The aggregation step is mathematically straightforward: w_global = Σ (n_k / n_total) * w_k for each client k.

Despite sharing a feature space, HFL systems frequently suffer from statistical heterogeneity in label distributions. A base station in a dense urban area may experience a radically different mix of traffic classes—streaming video versus IoT telemetry—compared to a rural macrocell.

• Label Distribution Skew: The prior probability P(y) of a specific traffic type or mobility pattern varies significantly per client.
• Concept Drift: The relationship between features and labels may also shift geographically due to different interference patterns.
• Mitigation often requires advanced optimization frameworks like FedProx or SCAFFOLD to correct for client drift during local training.

In production RAN intelligence, HFL is almost exclusively deployed in a cross-silo configuration. The clients are not millions of smartphones but a manageable number of O-RAN Central Units (O-CUs) or Distributed Units (O-DUs) with reliable power and high-bandwidth backhaul.

• Stateful Clients: Each base station node is always available and can maintain state across training rounds.
• Trusted Execution: Operators often combine HFL with Trusted Execution Environments (TEEs) on RAN Intelligent Controller (RIC) hardware to protect model integrity.
• This contrasts with cross-device HFL, which is impractical for latency-sensitive radio resource management tasks.

HFL protects data sovereignty by ensuring raw call traces and measurement reports never leave the base station. Only encrypted or differentially private gradient updates are transmitted to the central aggregation server.

• Secure Aggregation: A cryptographic protocol ensures the server can only compute the sum of updates, not inspect individual contributions.
• Differential Privacy: Calibrated Gaussian noise is added to gradients before transmission, bounding the influence of any single user's data.
• This architecture satisfies GDPR and telecom regulatory requirements by decoupling model training from data centralization.

The complementary topology to Horizontal Federated Learning where datasets share the same sample space but differ in their feature space. In a telecom context, this occurs when a mobile operator and a financial institution hold different attributes about the same set of subscribers. Training requires entity alignment to match records across silos before performing split neural network computation, often using privacy-preserving record linkage and secure multi-party computation to prevent feature leakage.

A hybrid topology applied when datasets differ in both sample space and feature space, with only partial overlap. This is common when two telecom operators in different regions serve entirely distinct user bases and collect slightly different network metrics. The technique leverages transfer learning to adapt a pre-trained base model to the target domain using limited overlapping data, minimizing the need for large labeled datasets in the target silo.

The institutional deployment model typically used for Horizontal Federated Learning in telecom. Involves a small number of reliable participants—such as regional base station clusters or operator data centers—with stable connectivity and substantial compute resources. Unlike cross-device settings, these silos are stateful and identified, enabling synchronous aggregation rounds and stronger security guarantees through Trusted Execution Environments.

The primary statistical obstacle in Horizontal Federated Learning. When different base stations serve demographically distinct user populations, the local data distributions diverge significantly from the global distribution. This statistical heterogeneity causes local models to drift apart during training, slowing global convergence. Mitigation strategies include:

• FedProx: Adds a proximal term to local objectives
• SCAFFOLD: Corrects client-drift using control variates
• Personalized FL: Balances global and local model performance

A cryptographic protocol essential for Horizontal Federated Learning deployments where the central server must compute the weighted average of model updates without inspecting any individual base station's gradient contributions. The server learns only the aggregated sum, protecting against gradient leakage attacks that could reconstruct user mobility patterns. Often combined with differential privacy to provide formal privacy guarantees against curious-but-honest aggregators.

An alternative to Horizontal Federated Learning where the neural network is partitioned between client and server. The base station processes initial layers on raw call detail records and transmits only intermediate activations (smashed data) rather than model updates. This reduces client-side computational burden and provides stronger raw data isolation, but introduces sequential training dependencies that can increase wall-clock training time compared to parallel federated averaging.