Membership inference attacks against synthetic health data

Area of Science:

• Health Informatics
• Data Privacy
• Machine Learning

Background:

• Synthetic data generation is a method for sharing health data while protecting patient privacy.
• Current synthetic data generators may not fully mitigate disclosure risks.
• Membership inference attacks pose an evolving threat to synthetic data privacy.

Purpose of the Study:

• To assess the disclosure risks of synthetic health data against membership inference attacks.
• To develop a framework for disclosure risk assessment from a data holder's perspective.
• To evaluate the effectiveness of state-of-the-art machine learning in enhancing these attacks.

Main Methods:

• Formulated membership inference as a disclosure risk assessment problem.
• Introduced a framework leveraging contrastive representation learning for enhanced attacks.
• Conducted experiments on synthetic health data derived from real-world resources (VUMC, All of Us).

Main Results:

• Partially synthetic data demonstrated high vulnerability to membership inference attacks.
• Fully synthetic data showed only marginal susceptibility to these attacks.
• The developed framework effectively assessed upper bounds of disclosure risk.

Conclusions:

• Existing synthetic data generators may not offer sufficient protection against advanced membership inference attacks.
• Partially synthetic data is particularly vulnerable, questioning its privacy-preserving claims.
• Fully synthetic data appears more robust, offering a potentially safer alternative for data sharing.