Federated learning is a machine learning technique proposed by researchers at Google, Inc. The concept and its core algorithm were formally published at the 2017 International Conference on Artificial Intelligence and Statistics (AISTATS 2017), in related papers. Communication-Efficient Learning of Deep Networks from Decentralized Data.

Federated learning is a privacy-preserving distributed machine learning method. Its core principle is to keep the original training data on the local device at all times, never uploading it to a central server. Instead, it collaboratively trains a shared global model by aggregating model update parameters calculated locally, thereby significantly reducing privacy leaks and security risks. To achieve this goal, the research team proposed the Federated Averaging (FedAvg) algorithm. Experiments have shown that this algorithm can not only stably handle the imbalanced and non-independent identically distributed (non-IID) data unique to mobile devices, but also significantly reduce the number of communication rounds required to train deep networks by 10 to 100 times, greatly overcoming the communication cost limitations in real-world applications.