Privacy-Preserving, Efficient, and Accurate Dimensionality Reduction

The swift evolution of artificial intelligence and big data has dramatically increased data volume and computational complexity, thereby considerably escalating data storage and processing costs. As a result, dimensionality reduction has become an essential phase in data pre-processing. Moreover, due to the privacy concerns associated with data gathered by various organizations, data sharing is restricted, introducing further challenges in data analysis and the training of machine learning models. Consequently, we introduce an innovative privacy-preserving dimensionality reduction scheme (PP-DR). PP-DR secures participant data using homomorphic encryption and eschews intricate bootstrapping tasks by employing secure interaction protocols, thereby efficiently performing joint dimensionality reduction on data shared among all participants. In contrast to current dimension reduction approaches employing homomorphic encryption, PP-DR achieves superior computational accuracy with an average error of only 1∗10−8. Additionally, it enhances computational efficiency by 30 to 200 times and reduces communication overhead by at least 70%. This study underscores the practical feasibility of secure multi-party collaborative dimension reduction using homomorphic encryption.