Hierarchical Aggregation (FedHier)

The foundational aggregation algorithm that constructs a global model by computing a weighted average of locally trained model updates.

• Mechanism: Server initializes a global model, distributes it to clients, clients train locally on private data, and the server averages the returned updates.
• Weighting: Updates are typically weighted proportionally to each client's local dataset size.
• FedHier Context: In hierarchical topologies, FedAvg is often the aggregation rule applied at both the edge server level (intermediate) and the cloud server level (final).

An intermediate computational node that sits between end-device clients and the central cloud server in a hierarchical topology.

• Function: Receives model updates from a regional subset of clients, performs a local aggregation step, and forwards the compressed result upstream.
• Benefits: Reduces wide-area network traffic, masks individual client contributions from the cloud, and enables geographic fault tolerance.
• Deployment: Often co-located with cellular base stations, hospital campus servers, or regional data centers.

A class of techniques designed to minimize the bandwidth overhead of transmitting model updates between aggregation tiers.

• Gradient Compression: Methods like sparsification (transmitting only top-k gradient elements) and quantization (reducing bit-width of updates) drastically reduce payload size.
• FedHier Synergy: Hierarchical aggregation naturally reduces communication by localizing traffic; combining it with compression creates a multiplicative reduction in core network load.
• Key Metric: Compression ratio vs. model accuracy degradation.

An aggregation scheme that updates the global model immediately upon receiving an update from any single client, eliminating the synchronization barrier.

• Straggler Mitigation: Prevents slow clients or network delays from blocking the entire training round.
• Temporal Inconsistency: Introduces challenges with staleness—the global model may have advanced while a slow client was training on an outdated version.
• Hierarchical Use: Edge servers can aggregate synchronously within their region while reporting asynchronously to the cloud, balancing efficiency and consistency.

A cryptographic protocol that allows a central server to compute the sum of encrypted client model updates without being able to inspect any individual contribution.

• Mechanism: Clients mask their updates with pairwise secrets that cancel out during summation, revealing only the aggregate.
• Hierarchical Extension: Multi-tier SecAgg protocols distribute trust across edge aggregators, ensuring no single node—including edge servers—can unmask individual updates.
• Privacy Guarantee: Provides information-theoretic privacy against honest-but-curious adversaries.

An aggregation framework that partitions clients into distinct clusters based on data distribution similarity, maintaining separate global models for each cluster.

• Heterogeneity Handling: Addresses non-IID data by grouping clients with similar clinical populations or imaging protocols before aggregation.
• Hierarchical Integration: Edge servers can perform cluster assignment locally, routing updates to the appropriate cluster-specific model maintained at the cloud tier.
• Algorithm: Iterative Federated Clustering Algorithm (IFCA) alternates between estimating cluster identities and optimizing cluster models.