Exploring the Trade-Off between Privacy and Predictive Power in Synthetic Data Generation

Data Scientists often come across a situation where the use of real data is often not recommended due to confidentiality or privacy concerns. The study offers a solution in the form of synthetic data that can be used in place of sensitive real data such that it follows the patterns and business constraints of real data while offering differential privacy and satisfactory predictive performance. The objective behind generating synthetic data is to achieve differential privacy. Differential privacy is a concept of sharing information about a dataset, such as interactions amongst features without giving away information about individuals in the dataset Increased differential privacy however comes at the cost of model performance. The purpose of this study is to explore different synthetic data generators and understand the trade-off between differential privacy offered by synthetic data and its impact on models' predictive power and identify the point at which there is a balance between the two. Our target audience is data scientists who will use such synthetic data in their analytics solutions to achieve significant business impact while maintaining data security and privacy of sensitive data.