Federated learning is a technique for training a shared machine learning model across many devices or organizations while keeping their raw data in place. Instead of sending personal or sensitive datasets to a central server, each participant trains the model locally and only transmits the resulting model updates, such as gradients or weights, back to a coordinating server. This setup reduces privacy risks and complies with regulations, because the actual data never leaves its original location. It is especially popular in situations where data is spread across smartphones, IoT devices, or organizational “data silos” and cannot be aggregated centrally due to legal, privacy, or logistical constraints. At its core, the concept follows the principle of “train where the data lives” and only transmit what the model has learned, not the data itself.
The training process occurs in coordinated cycles, referred to as federated learning rounds. It begins when a central server creates an initial global model. A subset of devices or organizations, called clients, is selected to take part in that round. Each client downloads the global model and trains it on its own local data for a short period. Afterwards, they send their trained model parameters or gradients back, often using encryption and privacy-preserving methods to ensure security. The server combines these contributions using a method like Federated Averaging, which calculates a weighted average based on each client’s dataset size, to form an updated global model. This updated model is then sent back to clients, and the loop repeats until the model reaches acceptable performance or meets a stopping condition.
Federated learning is valuable because it offers strong privacy benefits and enables organizations to collaborate without exposing proprietary or sensitive datasets. It also makes use of data generated at the network edge, improving personalization and reducing the need for massive centralized storage and bandwidth. However, it faces challenges such as handling data that is non-identically distributed across clients, coping with devices or organizations that may have varying computational resources or unreliable connectivity, and defending against risks like information leakage from model updates or malicious attempts to corrupt the model.
There are several flavors of federated learning. In horizontal federated learning, participants have similar feature spaces but different records, such as smartphones running the same app. Vertical federated learning occurs when participants possess different attributes about the same entities, for example when a bank and a retailer collaborate on shared customers. A more specialized approach, federated transfer learning, adapts pre-trained models across participants with limited data overlap. Deployments also differ by scale: cross-device scenarios involve many low-powered, often unreliable devices like phones or IoT sensors, whereas cross-silo settings involve a smaller number of stable, high-capacity organizations such as hospitals or financial institutions.
Despite the lack of centralized data sharing, privacy and security must be protected because model updates themselves can leak sensitive patterns. Countermeasures include secure aggregation, which uses cryptographic protocols so that the server can only view the combined update and not any individual contribution; differential privacy, which introduces noise to updates to limit the ability to infer personal data; and robust aggregation methods to defend against poisoned or manipulated updates.
Federated learning already appears in production systems, for example in mobile keyboards that improve word prediction models while keeping user typing data on the device, or in healthcare collaborations where hospitals jointly train models on imaging or diagnostics without sharing patient records. Building a simple federated system typically involves designing server and client roles, implementing aggregation logic, and integrating privacy safeguards before evaluating performance and scaling to deployment. As frameworks like Flower, TensorFlow Federated, or PySyft mature, it becomes increasingly feasible to design federated solutions that balance model utility, privacy guarantees, and system efficiency.