A local model in federated learning refers to the individual machine learning model that is trained on the data held on a specific device or node within a distributed environment. Instead of aggregating data from multiple sources onto a central server, federated learning allows each participating device to train its local model using local data. This approach minimizes data movement and helps protect user privacy since sensitive information remains on the user's device. Once training is complete, only the model updates—such as gradients or weights—are sent to a central server for aggregation, improving the global model while keeping raw data secure.
For example, consider a scenario where a company develops a predictive text model on mobile phones. Each phone collects unique typing habits and language preferences from its user. In traditional machine learning, all this data would need to be collected in one place for training, which could pose privacy concerns. With federated learning, each phone trains its own local model based on local data. This means that the personal typing patterns do not leave the device, yet they can contribute to improving a shared global model that better understands user behavior across many devices.
The strength of local models lies in their ability to adapt to the specific nuances of different users or environments. For instance, a local model on a user's device can better capture vocabulary specific to their region or interests, resulting in improved performance. Furthermore, local models can be periodically updated as new data becomes available, allowing them to evolve over time while still maintaining user privacy. Thus, local models are essential for federated learning, as they enable decentralized training while enhancing both security and personalization.