The growing adoption of machine learning across areas like healthcare and finance has raised serious concerns around data privacy. Traditional techniques such as differential privacy or federated learning often come with limitationswhether in accuracy, communication cost, or reliance on secure protocols. Fully Homomorphic Encryption (FHE) offers a promising alternative, enabling computation directly on encrypted data and making it possible to use data without ever seeing it in raw form.This paper explores how FHE can be integrated with machine learning workflows, from traditional models like linear regression and decision trees to complex deep learning architectures. We review mainstream FHE schemes and core technologies that make private ML feasible. In particular, we analyze the unique challenges of adapting different model types to FHE constraints and highlight real-world applications in medical imaging, financial models, and edge intelligence.However, critical bottlenecks remain. Large models still face efficiency issues, dynamic data settings are poorly supported, and the field lacks standardized benchmarks. Through this review, we outline key future research directions that can help transition FHE-based machine learning from theoretical promise to practical reality.
Qi Wang (Thu,) studied this question.