August 2024 Summaries

Gretel has announced the General Availability (GA) of Navigator Fine Tuning, a new addition to their suite of synthetic data solutions. This feature allows users to inject business and domain-specific knowledge into Navigator, their privacy-preserving compound AI system, by training it on relevant real-world datasets. Navigator Fine Tuning supports multiple tabular data modalities within a single model, including numerical, categorical, free text, and sequential (time-series) data. During the open preview period, hundreds of models were trained with thousands of minutes of API runtime from developers and enterprise customers. Navigator Fine Tuning is now Gretel's default model offered when selecting the 'start from scratch' blueprint on their console dashboard.

Recent concerns about model collapse have sparked debate in the AI landscape regarding the use of synthetic data for model development. While synthetic data offers immense potential, a study by Shumailov et al. raised questions about its impact on AI models. However, this extreme scenario of recursive training on purely synthetic data is not representative of real-world AI development practices. The combination of synthetic and real-world data can prevent model degradation, and thoughtful synthetic data generation rather than indiscriminate use is crucial for maximizing its potential benefits. Synthetic data has the potential to dramatically accelerate AI development across all sectors by filling critical data gaps, addressing biases, and creating more robust models.

This blog post provides a step-by-step guide to generating high-quality, privacy-safe synthetic patient data using Gretel's suite of tools. The process involves de-identifying personally identifiable information (PII) with Gretel Transform v2 and generating synthetic data with Navigator Fine-Tuning. This approach goes beyond simple PII removal, addressing the limitations of traditional anonymization techniques by creating new records not based on any single individual, providing robust protection against various privacy attacks and re-identification risks. The resulting synthetic dataset maintains the complex relationships and time-series nature of the original data, making it suitable for a wide range of healthcare analytics and machine learning tasks while preserving patient privacy.

Gretel introduces a new Data Privacy Protection Score that includes membership inference attack (MIA) and attribute inference attack (AIA) simulations, complementing the existing Privacy Configuration Score. The new metrics simulate adversarial attacks to measure privacy risk more effectively than traditional de-identification techniques. Gretel's Data Privacy Score helps businesses adhere to GDPR and similar regulations by integrating privacy-enhancing technologies such as de-identification, overfitting prevention, differential privacy, and continuous monitoring.