Training across many devices or silos without centralizing raw data, aggregating model updates instead.