Differentially Private Genomic Data Release For GWAS Reproducibility

Résumé : With the rapid development of technology in genome-related fields, researchers have proposed various approaches and algorithms in recent years. However, they rarely publish the genomic datasets they used in their works for others to reproduce and validate their methods, as sharing those data directly can lead to significant privacy risks (e.g., against inference attacks). To solve the problem and expedite cooperative scientific research, we propose a novel differentially private sharing mechanism for genomic datasets that protects the entire genomic dataset under differential privacy. To improve data utility of the GWAS statistics, we further develop a post-processing scheme that performs optimal transport (OT) on the empirical distributions of SNP values. The distributions are also achieved in a privacy-preserving manner. We evaluate our approach on several real genomic datasets and show in the experiments that it provides better protection against both genomic and machine learning-based membership inference attacks and offers higher GWAS utility than the baseline approaches.