Weighted Aggregation

The core mechanism scales each client's model update by a factor of n_k / n, where n_k is the local dataset size and n is the total aggregate sample count across all participating nodes. This statistical weighting assumes that models trained on more data produce more reliable gradient updates, naturally prioritizing institutions with larger patient cohorts.

Weighted Aggregation is the mathematical backbone of the Federated Averaging (FedAvg) algorithm. The server computes the new global model w_{t+1} as:

• w_{t+1} = Σ (n_k / n) * w^k_{t+1} This linear combination of local model weights ensures the global update is a statistically unbiased estimate of the true centralized gradient when data is IID.

In non-IID clinical settings, sample-size weighting can introduce bias. A large hospital with a narrow patient demographic may dominate the global model, skewing it away from rare diseases present in smaller, specialized clinics. This limitation motivates advanced strategies like FedProx and Agnostic Federated Learning (AFL).

To combat data imbalance, weighting can be extended beyond raw sample size to include label distribution or data quality metrics. Stratified weighting ensures that the global model does not overfit to a majority class from a large node, preserving diagnostic accuracy for underrepresented conditions across the network.

Weighted Aggregation operates on encrypted model updates, not raw patient data. When combined with Secure Aggregation (SecAgg) protocols, the server can compute the weighted sum without ever inspecting individual gradient contributions, maintaining strict HIPAA and GDPR compliance.

The aggregation step itself is computationally trivial compared to local training. However, the weighting coefficients n_k must be transmitted to the server. In cross-device FL, this metadata is minimal, but in cross-silo healthcare, institutions often negotiate dynamic weighting policies based on data contribution audits.

The foundational aggregation algorithm that constructs a global model by computing a weighted average of locally trained model updates. Each client's contribution is scaled proportionally to the size of its local dataset, giving more influence to nodes with larger clinical data volumes. This is the baseline against which all other aggregation strategies are measured.

Enhances FedAvg by adding a proximal term to local objective functions. This penalty term restricts local updates from diverging too far from the global model, stabilizing convergence across heterogeneous clinical devices and non-IID data distributions. Particularly effective when clients have variable computational capabilities.

A class of robust aggregation rules ensuring global model convergence even when a subset of nodes submits arbitrary, malicious, or corrupted updates. Critical for healthcare networks where compromised devices or adversarial actors could poison the collaborative model. Includes techniques like Krum and Trimmed Mean.

A cryptographic protocol allowing a central server to compute the sum of encrypted client updates without inspecting any individual contribution. Ensures that even the aggregator cannot access a single hospital's model parameters, providing a strong privacy guarantee complementary to differential privacy.

Injects calibrated statistical noise into the model update process, providing a formal mathematical guarantee that individual patient data cannot be inferred from the global model. Balances the privacy-utility tradeoff by bounding the influence of any single record on the aggregated output.

A multi-tier topology where edge servers perform intermediate model averaging on client updates before a central cloud server executes final global aggregation. Reduces latency and communication bottlenecks in large-scale healthcare networks spanning multiple hospital systems.