By Vishal Deshpande, Chief Data Analytics Officer
Using Profile-based Synthetic Data Generation to develop more responsible AI models
The demand for solutions leveraging artificial intelligence (AI) and machine learning (ML) continues to surge. Core to every AI or ML solution is a robust security model that protects data privacy. These more advanced data solutions demand a data security and privacy paradigm that evolves beyond approaches such as data suppression and data masking that, frankly, miss the mark when it comes to robust model security.
At Unissant, our teams help agencies identify the most secure and ethical pathways to implement AI/ML models. We avoid using personally identifiable information, public health information, or other confidential data, in production systems. Rather, we frequently create synthetic data that improves data privacy and security, enhances model performance, and accelerates AI development.
The idea of creating synthetic data is not new. However, traditional approaches have their limitations. Rule-based approaches to creating synthetic data only work for simple scenarios. Statistical approaches are good for general patterns, but they frequently fail to capture specific details. These approaches can be inflexible, limiting the variety and complexity of data they can generate. While data may appear statistically similar, data often lacks the nuances associated with real production data and can perpetuate bias.
To address these challenges, we developed a Profile-based Synthetic Generator. Profile-based synthetic data generation allows developers to construct and test AI/ML models without sacrificing precision.
What is Profile-based Synthetic Generation, you ask?
Let’s simplify the concept.
As a child, did you have an imaginary friend? Envision creating an imaginary friend with similar interests, likes, and experiences as your real friends. Profile-based synthetic data generation works much like this, only at scale. We give these fake people profiles filled with information like their age, gender, job, income, health records—whatever traits (or data fields) are relevant to the original data. The key—these profiles are real-world data, so they look and act like real people.
How does profile-based synthetic data make better models?
To answer that question, I’ll explain how data scientists “fake it” with profile-based synthetic data generation while making AI models.
At Unissant, we use profile-based synthetic data generation in the early stages of model development. By creating artificial datasets drawn from real-world data profiles, we generate extensive, dependable datasets. These realistic datasets support effective AI/ML model training and development while preserving anonymity. Importantly, this approach ensures that sensitive information remains protected. By adjusting the parameters we use to generate the data, we significantly reduce the impact of both implicit and explicit bias. In addition, we can develop multiple datasets for training and testing models. The result – better models that agencies can deploy with confidence.
Want to explore more?
If you’re intrigued by the idea of profile-based synthetic data generation, I invite you to read our white paper on the subject: Advancing Trustworthy AI/ML: Profile-based Synthetic Data Generation. To explore how profile-based synthetic data generation can support your programs objectives for responsible AI, reach out to me
