What is federated learning?

Federated learning is a machine learning approach where models are trained across multiple decentralized devices or servers without sharing raw data. Instead of centralizing data in one location, the model is sent to the devices (e.g., smartphones, IoT sensors, or local servers) where the data resides. Each device trains the model locally using its data, and only the model updates (like gradients or parameters) are sent back to a central server. The server aggregates these updates to improve the global model, which is then redistributed. This process repeats, allowing the model to learn from diverse data sources while keeping data private and localized.

A common example is a mobile keyboard app that learns personalized typing suggestions without uploading user text to the cloud. Each phone trains a language model locally, sends encrypted model updates, and the global model improves without exposing individual messages. In healthcare, hospitals could collaboratively train a diagnostic model without sharing sensitive patient records—each hospital trains on its data, and only model updates are combined. Federated learning is also used in IoT networks, where edge devices process sensor data locally, reducing bandwidth costs and latency compared to sending raw data to a central server. This approach is scalable for scenarios involving large, distributed datasets or privacy-sensitive domains.

Developers implementing federated learning face challenges like handling uneven data distribution (e.g., some devices may have biased or sparse data) and managing communication costs. Techniques like differential privacy or secure aggregation protocols can address privacy risks during update transmission. Frameworks like TensorFlow Federated or PyTorch’s Substra provide tools to simulate federated workflows, manage device-server communication, and handle aggregation logic. However, optimizing model performance requires balancing local training iterations with global synchronization frequency. For example, training too long on a device with skewed data might harm the global model’s generalization. Testing in heterogeneous environments and monitoring for convergence issues are critical steps in real-world deployments.