Federated Learning

Federated learning is a machine learning approach where a model is trained across multiple devices or organisations without any raw data leaving its source. Instead of collecting all data in one place, the algorithm sends the model to the data.

How federated learning works

1. A central server sends the current model to all participating devices or organisations
2. Each participant trains the model on their local data
3. Only the model updates (not the data) are sent back to the server
4. The server aggregates these updates into an improved global model
5. The cycle repeats until the model converges

The raw data never leaves each participant's control.

Why this matters for privacy

Traditional machine learning requires centralising data — a major barrier in regulated industries. Hospitals cannot share patient records. Banks cannot share customer transactions. Competing companies will not share proprietary data. Federated learning lets these organisations benefit from collective intelligence without exposing sensitive data.

Real-world applications

• Mobile keyboards — Google's Gboard learns from millions of users' typing patterns without uploading what they type
• Healthcare — hospitals train diagnostic models collaboratively without sharing patient data
• Finance — banks improve fraud detection by learning from industry-wide patterns without revealing customer information
• Manufacturing — factories share quality improvement insights without exposing trade secrets

Challenges

• Communication cost — sending model updates back and forth requires significant bandwidth
• Data heterogeneity — different participants have different types and distributions of data, which can make training difficult
• Security — while raw data stays local, model updates can potentially be reverse-engineered to infer information about the training data. Differential privacy techniques help mitigate this.
• Coordination — managing training across thousands of devices with varying availability, connectivity, and compute power is complex
• Free riders — some participants may benefit from the shared model without contributing useful updates