January 2021 Summaries

Growing up in Communist Romania, the author experienced a society where government surveillance was pervasive, influencing their lifelong concern about privacy. This concern is relevant today as digital footprints and data brokers reveal personal information without consent. Although advancements in computational power now allow rapid data analytics, many companies lack the necessary privacy safeguards, often leaving user data inadequately protected and accessible to too many individuals. Gretel.ai, inspired by the author's past and the need for privacy, offers a solution by enabling companies to access customer data without compromising individual privacy, demonstrating that privacy and innovation can coexist.

This post explores using synthetic data to balance a biased health dataset on Kaggle and improve overall model accuracy. The Heart Disease dataset published by the University of California Irvine is imbalanced, with male patient records accounting for 68% of the overall dataset and female patient records at only 32%. To reduce bias in the input data, synthetic female patient records were generated using Gretel.ai's open-source synthetic data library. The resulting augmented dataset was then run through ML algorithms on Kaggle to compare results against the original training set. In five out of six classification algorithms, accuracy increased when trained with the augmented dataset, achieving 96.7% overall accuracy for KNN (up from 88.5%) and 13% gains for the Decision Tree classifier.

This blog provides a step-by-step guide on using Gretel SDKs to create a fair, balanced, privacy preserving version of the 1994 US Census dataset. The process involves balancing underrepresented classes such as race, gender, and income bracket in the dataset. The Python notebook used for this purpose can be utilized for any imbalanced dataset. Gretel's SDK allows users to choose from two modes: "full" mode generates a complete synthetic dataset with representation bias removed, while "additive" mode only generates synthetic samples that remove bias when added to the original set. The blueprint also enables users to view existing categorical field distributions in the dataset and generate synthetic data for specific fields. Finally, users can save their new synthetic data or back onto a Gretel Project.

This post provides a practical guide to creating differentially private synthetic data using Python and TensorFlow. It demonstrates how to train a synthetic data model on the Netflix Prize dataset while protecting user identities through differential privacy techniques. The goal is to generate new data in the same format as the source data, with increased privacy guarantees and retaining statistical insights. The post discusses parameter tuning approaches for finding optimal privacy parameters and presents experiments using the gretel-synthetics library and TensorFlow-Privacy. It also explores optimizing learning rates, l2_norm_clip, and noise_multiplier to improve model accuracy while maintaining privacy guarantees. The final section encourages readers to experiment with generating synthetic datasets on their own data using the provided Jupyter notebook.

This post discusses building a data pipeline that automatically transforms datasets for safe use in development environments, ensuring customer privacy is maintained. The process involves using Gretel.ai's SDKs to auto-anonymize streaming data. The blog provides an open-source code blueprint detailing the steps to create such a pipeline. It covers labeling and discovery, rules evaluation, and record transformations. The resulting anonymized dataset can be safely pushed into pre-production environments without risk of leaking customer details.