SYNNER synthetic data generator framework.

Area of Science:

• Medical Informatics
• Data Science
• Privacy-Preserving Technologies

Background:

• Sharing sensitive medical data faces significant technical, regulatory, and privacy hurdles, including Health Insurance Portability and Accountability Act (HIPAA) compliance.
• Existing data anonymization techniques are often flawed, risking re-identification, while current synthetic data generation methods have limitations.

Purpose of the Study:

• Introduce SYNNER, a novel framework for generating synthetic data that preserves utility and ensures privacy.
• Overcome limitations of existing data anonymization and synthetic data generation methods.
• Facilitate secure data sharing and research in sensitive domains like healthcare.

Main Methods:

• Utilize knowledge graph embeddings to represent data in a k-dimensional space, capturing intricate relationships.
• Generate synthetic data by identifying nearest neighbors for each entity and replicating their characteristics to maintain statistical consistency.
• Introduce a novel evaluation protocol for differential privacy, including simulated adversarial attacks to assess re-identification risks.

Main Results:

• SYNNER preserves an average of 83.2% of signals from original datasets.
• Models trained on SYNNER data achieve a comparable average macro-F1 score of 74.4% in predictive tasks.
• The differential privacy evaluation protocol effectively assesses privacy standards and identifies potential reconstruction risks.

Conclusions:

• SYNNER offers a scalable and effective solution for generating statistically faithful synthetic data.
• The framework overcomes limitations of existing methods, providing a robust privacy-preserving approach.
• SYNNER advances research in sensitive areas by enabling secure and reliable synthetic data generation.