Federated Coreset Aggregation (FedCoreset)

Each client constructs a small, weighted subset of its local data that approximates the full dataset's loss gradient. This coreset is typically selected using greedy geodesic ascent or importance sampling to minimize the maximum error between the full gradient and the coreset gradient.

• Gradient Matching: Coreset points are chosen so their weighted gradient sum closely matches the full dataset gradient
• Bounded Error: Theoretical guarantees ensure the coreset introduces only ε-error in any direction
• Size Reduction: Typical coresets compress datasets by 10x to 100x while maintaining model quality

By transmitting only a synthetic coreset rather than raw data or full model updates, FedCoreset provides an inherent privacy barrier. The server never sees actual patient records, only a lossy compressed representation optimized for training.

• No Raw Data Exposure: Original clinical records never leave the local institution
• Reconstruction Resistance: Coreset points are weighted aggregates that resist direct inversion to original samples
• Compatible with DP: Coreset selection can be combined with differential privacy noise for formal guarantees

FedCoreset transmits a fixed-size data summary rather than full model parameters, making it ideal for bandwidth-constrained medical edge devices. The communication cost depends on coreset size, not model architecture.

• Constant Overhead: Communication cost is O(coreset_size × feature_dim) regardless of model complexity
• One-Shot Transmission: Clients send coresets once per round, eliminating iterative gradient exchanges
• Bandwidth Savings: Reduces communication by 50-90% compared to transmitting full model weights for large architectures

FedCoreset naturally handles statistical heterogeneity across clinical sites. Each client's coreset captures its local data distribution, and the server can apply weighted aggregation or clustered training on the collected coresets.

• Non-IID Resilience: Coresets preserve local distribution characteristics without homogenization
• Variable Coreset Sizes: Clients with more data can contribute larger coresets proportionally
• Stale Update Tolerance: Coresets remain valid even if global model drifts between rounds

Once coresets are collected, the server can train any model architecture on the aggregated coreset data. This decouples client computation from server optimization, enabling the use of adaptive optimizers and complex architectures centrally.

• Model Agnostic: Server can train CNNs, transformers, or any differentiable model on coreset data
• Full-Batch Optimization: Server can use L-BFGS or Adam on the combined coreset without client constraints
• Hyperparameter Tuning: Centralized training allows rapid experimentation without re-engaging clients

Clients perform lightweight coreset selection instead of full gradient-based training. This shifts the computational burden from resource-constrained edge devices to the well-provisioned central server.

• No Backpropagation: Clients avoid iterative forward-backward passes through deep networks
• One-Time Cost: Coreset construction is performed once per round, not over multiple local epochs
• Edge-Friendly: Enables participation from wearable devices and low-power imaging hardware that cannot run full SGD

A family of techniques that reduce communication overhead by transmitting sparsified or quantized gradient vectors. FedCoreset achieves a similar bandwidth reduction but operates in the data space rather than the gradient space.

• Top-K Sparsification: transmits only the largest gradient components
• QSGD: stochastically quantizes gradients to low precision
• Complements FedCoreset for multi-dimensional efficiency gains

An aggregation mechanism that injects calibrated statistical noise into the model update process. When combined with FedCoreset, the coreset selection itself can be performed under a differential privacy budget, providing formal guarantees that the selected representative samples do not leak individual patient information.

• Uses Gaussian or Laplacian noise mechanisms
• Privacy budget tracked via moments accountant
• Critical for HIPAA-compliant clinical deployments

The statistical challenge where local client datasets follow different probability distributions. FedCoreset addresses this by selecting a weighted subset that approximates the full local distribution, mitigating the representational bias that degrades standard FedAvg in heterogeneous clinical environments.

• Feature distribution skew: different hospitals image different populations
• Label distribution skew: rare diseases concentrated at specific sites
• Coreset selection can prioritize minority class representation

A cryptographic protocol enabling a server to compute the sum of encrypted client updates without inspecting individual contributions. When FedCoreset selects data points, the resulting model updates can be protected via SecAgg, ensuring that neither the raw data nor the coreset membership is exposed to the coordinating server.

• Uses secret sharing and pairwise masking
• Protects against honest-but-curious servers
• Adds communication overhead for cryptographic operations

Techniques that tailor global models to local patient populations. FedCoreset naturally supports personalization by allowing each client to maintain a site-specific coreset that captures its unique data distribution, enabling efficient local fine-tuning without storing the full dataset.

• FedBN: local batch normalization statistics
• pFedMe: Moreau envelope decomposition
• Coresets provide a compact, reusable local data summary