Privacy-preserving data sharing via probabilistic modeling.

Area of Science:

• Computer Science
• Statistics
• Epidemiology

Background:

• Differential privacy quantifies privacy loss from accessing sensitive data, with repeated access increasing loss.
• Releasing privacy-preserving synthetic data offers an alternative but requires effective design strategies.
• Current methods face challenges in balancing privacy guarantees with data utility.

Purpose of the Study:

• To propose a novel approach for private data release using probabilistic modeling.
• To address the challenge of designing high-quality synthetic data for research.
• To enable the creation of anonymized data twins for sensitive datasets.

Main Methods:

• Formulating private data release as a probabilistic modeling problem.
• Transforming synthetic data design into model selection.
• Incorporating prior knowledge to enhance synthetic data quality.

Main Results:

• Empirically demonstrated reliable reproduction of statistical discoveries from synthetic data in an epidemiological study.
• Showcased the effectiveness of probabilistic modeling in generating privacy-preserving synthetic data.
• Validated the approach for maintaining data utility and research reproducibility.

Conclusions:

• Probabilistic modeling offers a robust framework for generating high-quality, privacy-preserving synthetic data.
• The method facilitates the creation of anonymized data twins for sensitive datasets, supporting broad research applications.
• This approach is expected to significantly advance the field of private data sharing and analysis.