A global model in federated learning refers to a centralized machine learning model that is trained collaboratively by multiple devices or nodes without directly sharing their local data. Instead of sending raw data to a central server, each participant (often a mobile device or an edge device) trains a local version of the model using its own data. After this local training, each device sends only the model updates—such as weight changes or gradients—back to the central server. The server then aggregates these updates to improve the global model, which is subsequently distributed back to all participating devices for further training. This approach maintains data privacy, as the original data never leaves the user's device.
For example, consider a scenario where a company is developing a predictive text model for smartphones. Users' typing data remains sensitive and private, so instead of collecting this data in a central repository, the company uses federated learning. Each user’s phone trains a model based on their local typing habits. After local training, only model updates are sent to a server, where they are averaged with updates from other users’ phones. This aggregation process creates a more generalized global model that learns from the diverse typing patterns of all users while keeping individual data confidential.
By continuously repeating this training and aggregation process, the global model becomes more accurate over time, reflecting the insights derived from a wider audience without compromising user privacy. This collaborative approach not only fosters user trust but also enables the creation of robust models that can perform well across different contexts and user behaviors, ultimately leading to better and more personalized applications.