Federated Learning

The foundational principle where raw data never leaves its source device or institution. Training computation is pushed to the edge, and only model updates—such as gradient vectors or weight deltas—are transmitted. This directly addresses data residency requirements under regulations like GDPR and HIPAA, making it essential for genomic consortia where patient DNA cannot be pooled.

The canonical optimization algorithm that combines locally trained models into a global model. Each client trains on its local data for several epochs, then sends its updated weights to a central parameter server. The server computes a weighted average—typically proportional to each client's dataset size—to produce the next global model. This iterative process continues until convergence.

A cryptographic protocol ensuring the central server can only compute the sum of all client updates without inspecting any individual contribution. Using secret sharing and masking vectors, clients encrypt their updates such that pairwise masks cancel out during aggregation. If a client drops out, the server can still reconstruct the aggregate without compromising the privacy of remaining participants.

In real-world deployments, local datasets are rarely independently and identically distributed. A hospital specializing in rare diseases will have a vastly different genomic distribution than a general clinic. Federated learning frameworks must address this statistical heterogeneity through techniques like FedProx (adding a proximal term to local objectives) or personalized federated learning, which fine-tunes global models to each client's local distribution.

Transmitting full model updates over bandwidth-constrained networks is a primary bottleneck. Optimization strategies include:

• Gradient compression via sparsification (sending only the top-k gradients)
• Quantization (reducing 32-bit floats to 2-8 bits)
• Local updating (performing multiple local epochs before communicating) These techniques are critical for cross-device scenarios involving thousands of edge nodes.

Federated learning alone does not guarantee privacy—model updates can leak information through model inversion or membership inference attacks. To provide formal protection, differential privacy is integrated by clipping per-client updates and injecting calibrated Gaussian noise into the aggregated model. The privacy budget (ε, δ) quantifies the maximum information leakage, allowing institutions to provably bound re-identification risk for genomic data.

A data distribution characteristic where local datasets are not independently and identically distributed. In genomic federated learning, this arises when hospitals serve different populations with distinct allele frequencies, sequencing protocols, or disease prevalences.

• Label skew: Different hospitals have different disease distributions
• Feature skew: Different sequencing platforms produce varying read depths
• Requires specialized algorithms like FedProx to maintain convergence