A0548
Title: Federated statistical learning with differential privacy
Authors: Jaemu Heo - UNIST (Korea, South)
Jeonghun Kang - UNIST (Korea, South)
Taehwan Kim - UNIST (Korea, South)
Changgee Chang - Indiana University (United States) [presenting]
Abstract: While electronic health records (EHRs) offer great promises for advancing precision medicine, they suffer significant analytical challenges, which include the fact that it is often the case that patient-level data in EHRs cannot be shared across different institutions due to government regulations and/or institutional policies. A novel communication-efficient and privacy-preserving federated learning method is proposed to efficiently aggregate information from multiple datasets without exchanging individual patient-level data. Our new module INFEMBLER is presented, meaning an information assembler, which can extract and combine information carried in the perturbed MLE estimates from each remote database. The proposed approach allows proper statistical analyses from the linear model to various general models without sharing the raw patient-level datas. The proposed method's differential privacy properties and theoretical properties are investigated, and its performance is evaluated via simulations and real data analyses compared with several recently developed methods in the distributed statistical learning literature.
