Federated Learning: Definition and Examples

Federated Learning is a decentralized approach to training machine learning models. Unlike traditional methods that require centralizing all data on a single server, this technique allows training a model using data distributed across many devices (smartphones, hospitals, companies) without ever transferring them.

The process works in several steps: a global model is sent to each participant, who trains it locally on their own data. Only the model weight updates (the gradients) are sent back to the central server, which aggregates them to improve the global model. This cycle repeats until convergence. The best-known aggregation algorithm is FedAvg (Federated Averaging), proposed by Google in 2017.

The major benefit of Federated Learning lies in privacy protection. Sensitive data — whether medical records, personal messages, or financial data — never leaves the owner's device. This approach directly meets regulatory requirements such as the GDPR in Europe, while allowing to benefit from the power of large volumes of diverse data.

This technique is not without challenges: participants' data is often heterogeneous (non-IID), network connections can be unstable, and measures must be taken against poisoning attacks where a malicious participant tries to corrupt the global model. Complementary techniques such as differential privacy and homomorphic encryption are often combined with Federated Learning to strengthen privacy guarantees.