Global Model

The global model is mathematically synthesized by the aggregation server using algorithms like Federated Averaging (FedAvg). The server initializes a model architecture, distributes it to clients, and then computes a weighted average of the returned local model updates.

• Weighted Contribution: Each client's update is weighted by its local dataset size to prevent small datasets from skewing the global objective.
• Stateless Operation: The server holds no training data; it only processes the floating-point weight matrices received from clients.
• Convergence Target: The global model iteratively minimizes a global loss function that is the weighted sum of all local empirical losses.

The global model is designed to be data-agnostic. It never sees a single pixel of a medical scan or a single patient record. This architectural constraint is the primary mechanism for regulatory compliance.

• Gradient Isolation: Only encrypted or noise-masked model updates are transmitted, not raw data.
• Secure Aggregation (SecAgg): Multi-party computation protocols ensure the server can only decrypt the sum of updates, never an individual hospital's contribution.
• Differential Privacy (DP): The aggregator can inject calibrated noise into the global model to provide a mathematical guarantee against membership inference attacks.

A robust global model must generalize across Non-IID data distributions found in different hospitals. It learns features that are invariant across diverse scanner vendors, imaging protocols, and patient demographics.

• Mitigating Client Drift: Proximal optimization terms (e.g., FedProx) are used to constrain local updates from diverging too far from the global model during training.
• Universal Feature Representation: The global model distills common anatomical and pathological patterns while discarding site-specific biases.
• Cross-Site Validation: The ultimate test of a global model is its performance on a held-out test set from a completely unseen institution.

Once training converges, the global model is the canonical artifact deployed for inference. It is the consolidated intelligence of the entire consortium, often outperforming any single-hospital model.

• Centralized Deployment: The model can be deployed in a central cloud endpoint for tele-radiology services.
• Distributed Distribution: The final global model is pushed back to all participating edge nodes, upgrading every hospital's local diagnostic capability simultaneously.
• Version Control: The global model is checkpointed and versioned at each communication round, creating a full audit trail of its evolution.

The global model must be hardened against faulty or malicious client nodes. Robust Aggregation rules replace simple averaging to ensure a single compromised hospital cannot poison the collective intelligence.

• Byzantine Fault Tolerance: Algorithms like Krum or Trimmed Mean select or average only the most statistically consistent updates, ignoring outliers.
• Anomaly Detection: The aggregation server monitors the norm and direction of incoming gradients to flag suspicious updates for exclusion.
• Redundancy: The system relies on the statistical power of the honest majority to dilute the impact of adversarial noise.

The global model's update cycle is the primary bottleneck in federated systems. Transmitting full-precision weight matrices for large architectures like Vision Transformers can be bandwidth-prohibitive.

• Gradient Compression: Techniques like quantization (reducing 32-bit floats to 8-bit integers) and sparsification (sending only the top-k gradients) drastically reduce payload size.
• Federated Distillation: Instead of sharing weights, clients share only their model's output logits on a public, unlabeled dataset, decoupling model size from communication cost.
• Synchronous Rounds: The server must wait for straggler clients, making client selection and timeout management critical for operational efficiency.