Collaborative learning from distributed data with differentially private synthetic data.

Area of Science:

• Health Informatics
• Biomedical Research
• Data Privacy

Background:

• Collaborative learning is hindered by privacy concerns and the inability to pool sensitive data.
• Decentralized computation without central coordination poses challenges for joint analysis.
• This study explores using privacy-preserving synthetic data for collaborative learning on UK Biobank data.

Purpose of the Study:

• To evaluate the feasibility of combining synthetic data for collaborative learning.
• To assess the impact of data size, number of parties, and distribution shifts on learning outcomes.
• To determine if synthetic data sharing can overcome privacy and data access limitations in research.

Main Methods:

• Simulated multiple parties by splitting the UK Biobank cohort.
• Generated differentially private synthetic data for each simulated party.
• Applied Poisson regression analysis on combined synthetic data and compared with local data analysis.

Main Results:

• Collaborative learning with synthetic data yielded more accurate regression parameter estimates than using local data alone.
• Improvements were observed even with small, heterogeneous datasets.
• Increased participation of parties led to greater and more consistent improvements, up to a point.
• Synthetic data sharing particularly benefited analysis for underrepresented groups.

Conclusions:

• Sharing synthetic data is a viable strategy for privacy-preserving collaborative learning.
• This method enables learning from sensitive data without compromising privacy, even with limited or non-representative local datasets.
• Privacy-preserving collaborative learning methods can alleviate bottlenecks caused by inaccessible distributed sensitive data in biomedical research.