Cross-Silo Federated Learning

Unlike cross-device federated learning which scales to millions of unreliable edge devices, cross-silo FL involves a small, fixed set of institutional clients—typically 2 to 100 organizations. Each participant is:

• Always available: No random dropout or intermittent connectivity
• Stateful: Maintains persistent local datasets and training state across rounds
• Identifiable: Bound by legal contracts and service-level agreements This reliability eliminates the need for complex straggler mitigation and client selection strategies, simplifying orchestration.

Each silo possesses data-center-grade infrastructure rather than resource-constrained edge devices. This enables:

• Full local training: Each participant can train complete model epochs on their private data without offloading computation
• Larger batch sizes: Accelerating convergence compared to cross-device settings
• Complex local optimization: Techniques like FedProx with proximal terms can be applied without worrying about device battery life or memory limits This computational abundance shifts the bottleneck from client compute to communication efficiency between silos.

Participants are separated by institutional firewalls rather than device boundaries. This introduces unique requirements:

• Data sovereignty compliance: Model updates must respect jurisdictional regulations like GDPR, where data cannot cross geographic borders
• Contractual trust: Unlike anonymous cross-device clients, silos operate under formal data processing agreements
• Auditability: Each round of aggregation must be logged for regulatory inspection These boundaries make secure aggregation and differential privacy critical, even when participants are trusted entities.

Each institution's local dataset represents a non-IID distribution reflecting its specific user base or operational region. For example:

• Telecom Operator A: Urban 5G traffic patterns with high video streaming
• Telecom Operator B: Rural coverage with IoT sensor telemetry This statistical heterogeneity causes local optima to diverge, challenging standard Federated Averaging (FedAvg). Mitigation strategies include:
• Proximal terms in local objectives to anchor updates near the global model
• Knowledge distillation to transfer soft labels rather than raw weights

Cross-silo FL supports two fundamental data partitioning schemes:

• Horizontal Federated Learning: Silos share the same feature space (e.g., call detail records) but serve different user populations. Common in telecom where operators collect identical metrics for distinct subscribers
• Vertical Federated Learning: Silos hold different attributes about overlapping entities. For instance, a bank and a telecom operator may share customers but possess complementary features—financial history vs. mobility patterns—enabling joint fraud detection without exposing raw columns

With abundant local compute, the limiting factor becomes bandwidth between institutional data centers. Each round requires transmitting full model updates—potentially gigabytes for large neural networks. Optimization techniques include:

• Gradient compression: Quantization and sparsification reduce update payload sizes by up to 300x
• Reduced synchronization: Increasing local epochs per round minimizes communication frequency
• Secure aggregation overhead: Cryptographic protocols like SMPC add communication rounds, requiring careful protocol design to balance privacy and throughput

A cryptographic protocol ensuring the central server learns only the sum of model updates, not individual contributions. This prevents the aggregator from inspecting a telecom operator's proprietary gradient information. It is typically implemented using Secure Multi-Party Computation (SMPC) to mask updates, guaranteeing that no single participant's local data or model can be reverse-engineered during the global model assembly process.

A mathematical framework that provides a quantifiable privacy guarantee by injecting calibrated statistical noise into model updates. In a cross-silo context, this prevents membership inference attacks that could reveal if a specific subscriber's data was used in training. The privacy loss is controlled by the epsilon (ε) parameter, where a lower epsilon enforces a stronger guarantee, ensuring compliance with data sovereignty regulations.

A cryptographic primitive allowing computations directly on encrypted ciphertext. In federated learning, a telecom operator can encrypt its local gradient, the aggregator performs the weighted averaging on the encrypted data, and only the final encrypted global model is returned. The aggregator never possesses the decryption key, providing a strong hardware-agnostic guarantee against data exposure during the aggregation phase.

The condition where local datasets across silos are non-IID (non-Independently and Identically Distributed). For example, a base station in a business district sees different traffic patterns than one in a residential area. This violates standard optimization assumptions and causes model divergence. Cross-silo frameworks like FedProx address this by adding a proximal term to local objectives, stabilizing convergence despite skewed local data distributions.

A topology where datasets share the same sample space but different feature spaces. For instance, a mobile network operator holds call detail records while a financial institution holds credit scores for the same subscribers. Entity alignment techniques are used to match records without exposing raw IDs, allowing the collaborative training of a model that leverages both feature sets without either party revealing its proprietary columns.

The resilience property enabling correct consensus even when a subset of institutional nodes exhibits arbitrary or malicious behavior. In a cross-silo consortium, a compromised operator might send poisoned updates to corrupt the global model. Robust aggregation rules, such as Krum or trimmed mean, filter out these anomalous gradients, ensuring the global model remains reliable despite adversarial participants.