Federated Learning

The foundational principle: raw training data never leaves the edge device or local server. The global model is dispatched to each client node, where training occurs directly on the local dataset. This enforces data minimization and purpose limitation by design, as the central server never has access to the raw, sensitive information. This is a critical control for compliance with regulations like GDPR and the EU AI Act.

The core communication protocol. After local training, each client transmits only its model weight updates (gradients or parameters) back to a central coordinating server. The server then applies an aggregation algorithm, most commonly Federated Averaging (FedAvg) , to mathematically fuse these disparate updates into a single, improved global model. This process distills collective intelligence without exposing individual data points.

A primary technical challenge. Unlike a curated central database, data distributed across client devices is non-Independent and Identically Distributed (non-IID) . A user's photo library, typing history, or financial transactions are highly personal and do not represent the overall population. Federated learning algorithms must be robust to this statistical heterogeneity to prevent local biases from degrading the global model's performance.

A cryptographic enhancement ensuring the central server can only decrypt the sum of model updates from a group, not any individual client's contribution. Using Secure Multi-Party Computation (SMPC) , clients mask their updates with shared secrets. The masks cancel out upon summation, revealing only the aggregate result. This provides a robust defense against a curious or compromised aggregator attempting to infer private data from individual weight updates.

A mathematical noise-injection technique applied to model updates before transmission. By clipping and adding calibrated statistical noise to each client's gradient, a privacy budget (epsilon) is enforced. This provides a formal, quantifiable guarantee that the global model's output is indistinguishable from one trained without any single client's data, mitigating membership inference attacks and providing a strong layer of privacy-preserving machine learning.

Two distinct operational topologies. Cross-device FL involves millions of unreliable, low-power edge devices (e.g., smartphones, IoT sensors) with intermittent connectivity. Cross-silo FL involves a small, stable consortium of institutional data silos (e.g., hospitals, banks) with reliable compute resources. The latter is often used for healthcare federated learning to train diagnostic models across multiple hospitals without centralizing patient health information.

The foundational privacy principle that federated learning operationalizes. Raw data never leaves the edge device; only abstract mathematical updates are exchanged. This directly satisfies the GDPR requirement to limit processing to what is adequate, relevant, and limited to the specified purpose.

• Eliminates centralized data lakes as attack surfaces
• Reduces compliance scope for data breach notification
• Aligns with privacy by design architecture mandates
• Contrasts sharply with traditional data warehouse approaches

The architectural practice of logically or physically segregating datasets to prevent cross-contamination between models. In federated systems, this isolation is inherent: each client's data remains in its own siloed environment, and the orchestrator never accesses raw records.

• Prevents function creep where data collected for one purpose trains unrelated models
• Enforces purpose limitation at the infrastructure level
• Uses namespace isolation and strict IAM policies
• Critical for multi-tenant enterprise federated deployments