Fully Homomorphic Encryption and Artificial Intelligence: A Review of Architectures, Optimizations, and Distributed Privacy
Abstract
Fully homomorphic encryption (FHE) enables arbitrary mathematical evaluation directly over encrypted ciphertexts without intermediate decryption, establishing a foundational cryptographic primitive for cloud computation and secure machine learning. However, practical real-world adoption has historically been constrained by computational latency, ciphertext expansion, and memory bandwidth bottlenecks. To address these bottlenecks, this review examines state-of-the-art advancements across four key domains: 1) hardware acceleration via Graphics Processing Units, Field-Programmable Gate Arrays, and Application-Specific Integrated Circuits to reduce latency; 2) algorithmic optimizations like Single Instruction Multiple Data batching to maximize throughput; 3) Privacy-Preserving Machine Learning techniques for evaluating non-linear functions; and 4) Federated Learning integration using secure aggregation. This review explicitly demarcates the operational boundaries of Privacy-Preserving Machine Learning (PPML): while encrypted inference evaluates pre-trained models over encrypted inputs with static plaintext weights at amortized latency, fully homomorphic training demands dynamic encrypted weight-activation multiplications, multi-layer gradient backpropagation, and recurrent bootstrapping. Throughput achieved in encrypted inference does not directly transfer to end-to-end encrypted training because of multiplicative depth limits and recurrent bootstrapping. Together, these algorithmic and architectural developments establish practical pathways for deploying privacy-preserving machine learning systems under rigorous cryptographic guarantees.