Privacy-Preserving Collaborative Population Stratification with Dynamic Algorithm and Hyperparameter Selection

Area of Science:

• Genomics
• Computer Science
• Privacy-Enhancing Technologies

Background:

• Population stratification is crucial for genetic studies.
• Existing methods often lack privacy guarantees for collaborative analysis.
• Differential privacy (DP) offers robust privacy but can impact utility.

Purpose of the Study:

• To introduce a novel privacy-preserving selection layer for collaborative population stratification.
• To enable automatic selection of optimal analysis pipelines under epsilon-local differential privacy (LDP).
• To balance privacy, utility, and computational efficiency in federated genomic analyses.

Main Methods:

• Developed a framework offering three DP pipelines: PCA→Noise, Noise→PCA, and Noise-Only.
• Implemented an honest-but-curious server to aggregate DP shares for automated algorithm and parameter selection (K-Means, GMM, Hierarchical, K).
• Evaluated pipeline performance using internal metrics (Silhouette, Calinski-Harabasz, Davies-Bouldin) on the openSNP dataset.

Main Results:

• PCA-augmented pipelines demonstrated higher utility and significantly lower communication/runtime compared to Noise-Only.
• The automatically recommended configuration consistently outperformed fixed baseline methods.
• PCA-based pipelines showed markedly lower membership-inference attack power across various privacy budgets (epsilon).

Conclusions:

• The proposed privacy-preserving selection layer effectively automates pipeline selection for collaborative population stratification.
• PCA-integrated DP pipelines offer a superior balance of utility, efficiency, and privacy.
• Future work will explore extensions to multi-site collaborations.