HEJet: A Framework for Efficient Machine Learning Inference with Homomorphic Encryption
David Monschein, Oliver P. Waldhorst · 2024
The increasing adoption of machine learning (ML)-based services has presented challenges in processing sensitive data while ensuring privacy and confidentiality. Homomorphic encryption offers a promising solution by enabling computations on encrypted data. However, applying homomorphic encryption in ML faces challenges regarding efficient structuring, arrangement, and execution of numerical operations. In this paper, we present HEJet: a framework that enables efficient and user-friendly application of neural networks with homomorphic encryption. Our framework maps sequences of numerical computations to an optimized set of instructions that are processed by compilers for homomorphic encryption. Consequently, HEJet provides user-friendly interfaces to utilize advanced neural network structures with homomorphic encryption. Evaluation results on the MNIST dataset highlight its usability and show a significant speedup in inferences between 3% and 48% compared to existing approaches. Additionally, HEJet maintains accuracy levels close to those observed on raw data.