Integrating differential privacy into federated multi-task learning algorithms in dsMTL.

Area of Science:

• Computational biology
• Bioinformatics
• Statistical learning

Background:

• Multi-task learning (MTL) leverages shared information for simultaneous regression or classification tasks.
• The dsMTL R package facilitates federated MTL using the DataSHIELD platform for distributed, sensitive data analysis.
• Existing dsMTL security is insufficient against membership inference attacks targeting training data inclusion.

Purpose of the Study:

• To enhance the privacy-preserving capabilities of dsMTL against membership inference attacks.
• To integrate differential privacy into dsMTL for protecting individual-level data in federated learning.
• To balance privacy protection with model performance in federated MTL.

Main Methods:

• Implemented differential privacy using the Laplace mechanism as an optional feature in dsMTL.
• Utilized simulation studies and a case study on schizophrenia patient gene expression data for validation.
• Focused on obscuring individual characteristics while preserving group-level differences.

Main Results:

• Successfully integrated differential privacy into the dsMTL framework.
• Demonstrated the effectiveness of differential privacy in protecting against membership inference attacks.
• Validated the approach through simulations and a real-world gene expression data analysis.

Conclusions:

• The novel differential privacy feature in dsMTL significantly enhances data security for federated MTL.
• Achieving an optimal balance between privacy and model performance is critical for practical application.
• dsMTL with differential privacy offers a robust solution for privacy-preserving analysis of distributed sensitive data.