Regression on Encrypted Datasets Using Fully Homomorphic Encryption

This technical report discusses the application of Fully Homomorphic Encryption (FHE) in the context of training linear regression models on encrypted datasets. It specifically focuses on the CKKS scheme, which allows operations on encrypted floating-point numbers, making it suitable for real-world data analytics. The authors implement a gradient descent algorithm using FHE primitives to train a linear regression model and perform inference on data obtained from the Private Set Intersection (PSI) of two datasets. The report outlines the significance of data privacy in a data-driven world and the advancements in FHE as a solution to privacy concerns. It details the architecture of the implementation, emphasizing the use of hardware acceleration to enhance performance. The findings indicate a significant speedup in the encrypted gradient descent algorithm, demonstrating the potential of FHE for privacy-preserving machine learning applications. The report also reviews related works and discusses the challenges and methodologies associated with training regression models on encrypted data.