Privacy Risk Assessment for Synthetic Longitudinal Health Data.

Area of Science:

• Data privacy
• Synthetic data generation
• Epidemiology

Background:

• Synthetic data generation methods offer a modern approach to data privacy, often claiming superior utility-privacy trade-offs compared to traditional anonymization.
• Deep learning models can generate useful synthetic datasets, but evaluating their privacy implications, especially concerning data protection guidelines, remains difficult.

Purpose of the Study:

• To evaluate the privacy implications of synthetic data generated for an epidemiological study.
• To assess the effectiveness of the Anonymeter framework in quantifying privacy risks in synthetic datasets.
• To identify specific vulnerabilities, particularly concerning outliers, within synthetic data.

Main Methods:

• Application of the Anonymeter privacy risk quantification framework.
• Analysis of synthetic data generated from the DONALD cohort study (1312 participants, 16 time points).
• Focus on privacy risks including singling out, linkability, and attribute inference, with an emphasis on outlier vulnerability.

Main Results:

• Privacy scores varied significantly across different attack types.
• The study identified specific vulnerabilities related to outliers in the synthetic data.

Conclusions:

• Privacy risk assessment for synthetic data is an ongoing challenge.
• Implementation and interpretation of privacy risk assessment results present difficulties.
• Further research is needed to develop robust methods for synthetic data privacy risk assessment.