Predicting Genetic Disorder on Encrypted Data with Concrete ML Homomorphic Encryption Library

Abstract

This paper investigates the use of Fully Homomorphic Encryption (FHE) through the ConcreteML library to ensure data privacy in machine learning applications. The research focuses on predicting genetic disorders using a Gradient Boosted Tree model trained on a dataset of 22,000 genomic samples. ConcreteML enables encrypted inference without exposing sensitive patient data during cloud-based processing. The results show that FHE-enabled models maintain high accuracy, comparable to their plaintext counterparts, with acceptable performance trade-offs. The thesis also discusses key technical challenges, including quantization, model compilation, and managing computational overhead. Ultimately, this work demonstrates the viability of secure, privacy-preserving machine learning in genomics using modern FHE tools.