Random Forest in Federated Learning Setting

Federated learning has become popular nowadays since being started by Google in developing their Google Keyboard solution. Original federated learning proposed how to secure aggregating the gradient from the local training process of joining mobile devices. However, other machine mechanisms without gradient updating, such as random forest, are not supported. In this work, we propose PriForest, a new approach for building random forest in the federated learning setting. This framework covers all random forest processes, including building the forest, updating the forest, and using the forest for prediction. For providing privacy reserving, we apply a differential-privacy scheme while making bootstrapping sets, which modifies the value of existing records or appends new records before constructing trees. The experiment result shows that our approach is applicable with an affordable noise rate, which controls the sensitivity of the noise adding method.