Fidelity-agnostic synthetic data generation improves utility while retaining privacy.

Area of Science:

• Data Science
• Privacy-Preserving Technologies
• Machine Learning

Background:

• Synthetic data generation is crucial for sharing human subject data while maintaining privacy.
• Current methods prioritize statistical similarity, potentially including irrelevant information that compromises privacy.
• Balancing data utility and privacy is a key challenge in synthetic data generation.

Purpose of the Study:

• To propose a novel synthetic data generation method that prioritizes relevant features for specific applications.
• To enhance privacy protection by omitting irrelevant information from synthetic datasets.
• To demonstrate improved performance in prediction tasks using the proposed synthetic data.

Main Methods:

• Developed a fidelity-agnostic synthetic data generation approach.
• Utilized a neural network to extract relevant features for the dataset's intended use.
• Generated synthetic features and decoded them to mimic real data characteristics.

Main Results:

• The proposed synthetic data improved performance in prediction tasks.
• The method demonstrated enhanced privacy protection compared to existing state-of-the-art techniques.
• Fidelity-agnostic synthetic data successfully balanced utility and privacy.

Conclusions:

• Synthetic data generation can be optimized by focusing on task-relevant features, not just overall statistical similarity.
• The fidelity-agnostic method offers a promising approach for generating high-utility, privacy-preserving synthetic datasets.
• This technique has broad applicability across scientific domains requiring sensitive data analysis.