Abstract
The growing complexity of artificial intelligence (AI) models requires significant computing power to train them, posing a challenge for individuals and organizations without the necessary resources. Cloud-based training services can offer a viable solution, but they require sharing sensitive data with unvetted actors, presenting a risk to data privacy. Homomorphic encryption allows computation on encrypted data but adds computational complexity. To mitigate this problem, we explore using distributed training in combination with homomorphic encryption to parallelize the process of training on encrypted data. Using CKKS, a homomorphic encryption scheme that performs operations on real numbers, we develop a framework that trains comparably accurate AI models in less time than other homomorphically encrypted training solutions. Our experiments demonstrate reduced total runtime for homomorphically encrypted model training while maintaining competitive classification accuracy for the MNIST handwritten digits dataset, a well-known benchmarking dataset for machine learning. Our framework brings homomorphic encryption closer to a practical data privacy solution for small stakeholders who cannot compromise on security.