Generalized genomic data sharing for differentially private federated learning.

Area of Science:

• Genomics
• Machine Learning
• Data Privacy

Background:

• Federated Learning (FL) enables insights from distributed datasets, but genomic data requires enhanced security due to its sensitivity.
• Sharing sensitive genomic data for Machine Learning (ML) necessitates robust privacy mechanisms.

Purpose of the Study:

• To propose a generalized, privacy-preserving method for sharing gene expression data for Federated Learning.
• To reduce data dimensionality for efficient privacy budgeting using an exponential mechanism.

Main Methods:

• Utilized a differentially private mechanism for secure gene expression data sharing.
• Employed an exponential mechanism to create private histograms from numeric expression data, reducing dimensionality.
• Applied the method to federated machine learning settings with multiple data owners.

Main Results:

• The proposed solution ranked third out of 55 teams in the iDash 2020 genomic data security competition.
• Experimental results demonstrated model training in approximately 8 seconds with an AUC of 0.89 at a privacy budget of 5.
• The method proved effective across different machine learning algorithms and settings.

Conclusions:

• The outlined method effectively shares gene expression data for Federated Learning while preserving privacy.
• The approach demonstrates efficacy and scalability for genomic datasets and ML algorithms, with potential for broader applications.