FEDSET: Federated Random Forest Based on Differential Privacy

The rapid development of the federated machine learning paradigm has broken down the data barriers in technology between different organizations and individuals, allowing data that was previously difficult to analyze collaboratively to be used to unimaginable value. Among them, federated random forest has gradually gained popularity among collaborative analysis users for its good adaptability to structured data. However, the contradiction between the demand for collaborative analysis services and users’ awareness of privacy-preserving is becoming increasingly prominent and has become a bottleneck hindering the healthy and orderly development of the big data industry. To this end, this paper comprehensively analyzes the privacy preservation and usability requirements faced by federated random forest, and investigates a differential privacy-based federated random forest approach FEDEST. Users design privacy budgets locally, build CART decision trees with noise through a selective noise addition mechanism, and then upload them to the server for aggregation into a random forest, and optimize the random forest model through multiple rounds of iterations. FEDEST can improve the accuracy of business analysis models while safeguarding the privacy information of collaborative analysis participants. Experiments on the Adult and BRFSS datasets show that FEDEST has the highest classification accuracy of 91% and 87%, respectively, which is close to the classification accuracy of non-federated forest.