International Business Machines Corporation
Anonymizing data for preserving privacy during use for federated machine learning

Abstract:

A computer-implemented method for training a global federated learning model using an aggregator server includes training multiple local models at respective local nodes. Each local node selects a set of attributes from its training dataset for training its local model. Each local node generates an anonymized training dataset by using a syntactic anonymization method, and by selecting quasi-identifying attributes from training attributes, and generalizing the quasi-identifying attributes using a syntactic algorithm. Further, each local node computes a syntactic mapping based on equivalence classes produced in the anonymized training dataset. The aggregator server computes a union of mappings received from all the local nodes. Further, federated learning includes training the global federated learning model by iteratively sending, by the local nodes to the aggregator server, parameter updates computed over the local models. The aggregator server aggregates the received parameter updates, and sends the aggregated parameters to the local nodes.