FedAvg

The foundational federated learning algorithm that aggregates local model updates from decentralized devices to train a global model without moving raw data.

Introduced by Google researchers in 2016 (McMahan et al.), Federated Averaging (FedAvg) solves the privacy-utility tradeoff by keeping data on-device. The process is surgical: a central server broadcasts the current model to a subset of clients (like Android phones or IoT sensors); these clients perform local Stochastic Gradient Descent (SGD) on their private datasets; finally, the server collects these weights and computes a weighted average to update the master model. This approach slashes communication costs by 10x to 100x compared to synchronized gradient sharing and serves as the backbone for production systems like Gboard's predictive text.

https://arxiv.org/abs/1602.05629