Abstract: Embodiments of the present disclosure provide hierarchical, differential privacy enhancements to federated, machine learning. Local machine learning models may be generated and/or trained by data owners participating in the federated learning framework based on their respective data sets. Noise corresponding to and satisfying a first privacy loss requirement are introduced to the data owners' respective data sets, and noise corresponding to and satisfying a first privacy loss requirement are introduced to the local models generated and/or trained by the data owners. The data owners transmit model data corresponding to their respective local models to a coordinator, which in turn aggregates the data owners' model data. After introducing noise corresponding to and satisfying a third privacy loss requirement to the aggregated model data, the coordinator transmits the aggregated model data to the data owners to facilitate updating and/or re-training on their respective machine learning models.

Abstract: Embodiments of the present disclosure provide hierarchical, differential privacy enhancements to federated, machine learning. Local machine learning models may be generated and/or trained by data owners participating in the federated learning framework based on their respective data sets. Noise corresponding to and satisfying a first privacy loss requirement are introduced to the data owners' respective data sets, and noise corresponding to and satisfying a first privacy loss requirement are introduced to the local models generated and/or trained by the data owners. The data owners transmit model data corresponding to their respective local models to a coordinator, which in turn aggregates the data owners' model data. After introducing noise corresponding to and satisfying a third privacy loss requirement to the aggregated model data, the coordinator transmits the aggregated model data to the data owners to facilitate updating and/or re-training on their respective machine learning models.