Privacy-preserving verification of preprocessing in federated learning for genomic data.

Area of Science:

• Genomics
• Computational Biology
• Privacy-Preserving Technologies

Background:

• Federated learning (FL) enables collaborative genomic studies without centralizing raw data.
• Ensuring consistent data preprocessing across sites is crucial for reliable results.
• Protecting participant privacy is paramount in genomic data analysis.

Purpose of the Study:

• To develop and validate a method for verifying identical preprocessing pipelines in federated genomic studies.
• To ensure data integrity and reproducibility without compromising raw genotype privacy.

Main Methods:

• Institutions applied local differential privacy (LDP) to perturb genomic data slices.
• A RandomForest classifier was trained locally, and LIME explanation vectors were transmitted.
• A central server simulated preprocessing combinations and trained a classifier to detect site configurations.

Main Results:

• Centralized simulations achieved 80% accuracy in verifying preprocessing across 15 configurations.
• Membership-inference attack power was maintained below 0.05 at ε=3.
• Distributed FL experiments reached 70% accuracy in binary compatibility detection.

Conclusions:

• A single differentially private explanation vector serves as an auditable preprocessing fingerprint.
• The framework demonstrates feasible automated preprocessing verification in federated genomic consortia.
• This approach maintains participant privacy while ensuring data standardization.