Role of Homomorphic Encryption in Privacy-Preserving Machine Learning

Abstract

Homomorphic encryption (HE) has emerged as a pivotal cryptographic technique for enabling end-to-end privacy in machine learning workflows. By allowing arbitrary computations on encrypted data without exposing plaintext, HE addresses stringent privacy requirements across domains such as healthcare, finance, and telecommunications. This manuscript deepens the exploration of HE’s role in privacy-preserving machine learning (PPML) by expanding upon algorithmic foundations, practical implementations, and performance considerations. We provide an enriched theoretical overview of partially homomorphic (PHE), somewhat homomorphic (SHE), and fully homomorphic encryption (FHE) schemes, alongside a detailed comparison of their arithmetic capabilities, noise management strategies, and security parameters. A comprehensive simulation study on a logistic regression classifier trained with the UCI Heart Disease dataset is presented, contrasting plaintext, Paillier-based PHE, and CKKS-based FHE modes. Our extended statistical analysis quantifies not only model accuracy and computational latency but also communication overhead, ciphertext size inflation, and resource utilization.

Simulation research elucidates end-to-end encrypted workflows, highlighting batching strategies, polynomial activation approximations, and bootstrapping schedules. Results reveal that FHE can achieve confidentiality across training and inference with minimal accuracy loss (<3%), albeit with 8–10× training time overhead and 5–15× inference latency. We discuss advanced optimizations—including hybrid HE-MPC pipelines, hardware accelerators, and domain-specific parameter tuning—to narrow performance gaps. Finally, we outline future research directions in scalable HE libraries, federated learning integration, and adaptive noise budgeting, offering a roadmap toward practical, efficient PPML systems.