Communication Round

The round begins with the aggregation server broadcasting the current global model weights to a selected subset of participating client nodes. This step ensures all clients start local training from a synchronized state. The selection strategy can be random or based on specific eligibility criteria like device charging status or network connectivity.

Each selected client performs stochastic gradient descent (SGD) on its private, local dataset for a fixed number of epochs. This step generates a set of weight updates (gradients) that capture the unique statistical patterns of the local data without exposing the raw data itself. The computational burden is distributed, preserving data locality.

Clients encrypt their computed weight updates using protocols like Secure Aggregation (SecAgg) or Differential Privacy (DP) before transmission. This cryptographic layer ensures the central server cannot inspect any single institution's contribution in plaintext, mitigating the risk of model inversion attacks and maintaining strict regulatory compliance.

The server collects the encrypted updates and applies an aggregation algorithm, most commonly Federated Averaging (FedAvg). It computes a weighted average of the updates to produce a new, improved global model. Robust aggregation variants like Krum can be used to defend against Byzantine failures or poisoned updates.

The cycle repeats for hundreds or thousands of rounds until the global model's performance converges to a satisfactory accuracy threshold. The total number of rounds required is heavily influenced by the degree of statistical heterogeneity (Non-IID data) across the client silos, with more heterogeneous data often requiring more communication rounds.

A critical bottleneck in this cycle is the bandwidth required to transmit large model weights. Techniques like gradient compression (sparsification and quantization) and Federated Distillation are often applied to reduce the payload size, making the round viable for cross-device settings with limited network capacity.

The foundational aggregation algorithm that powers most communication rounds. After local training, clients send their model weight updates to the server, which computes a weighted average to produce the new global model. This simple yet effective mechanism allows collaborative learning without raw data exchange.

A primary challenge occurring during the local training phase of a round. When client datasets are Non-IID, local models diverge from the optimal global objective. This drift slows convergence and can degrade the final model's performance, requiring proximal terms or variance reduction techniques.

A cryptographic protocol that secures the server-side step of the round. It allows the server to compute the sum of encrypted model updates without inspecting any individual client's contribution in plaintext. This ensures that even the aggregator cannot steal proprietary institutional knowledge.

A communication efficiency technique applied before transmitting updates. Methods like quantization (reducing bit precision) and sparsification (sending only significant gradients) drastically reduce the bandwidth required per round, which is critical for hospitals with limited network infrastructure.

Defense mechanisms for the aggregation step against Byzantine failures. Rules like Krum or Trimmed Mean filter out malicious or corrupted updates from compromised clients, preventing a single bad actor from poisoning the global model during a communication round.

A mathematical guarantee often integrated into the update step. By clipping and adding calibrated Gaussian noise to model updates before transmission, DP ensures that the server cannot infer whether a specific patient's data was included in the local training batch.