Data Collaboration Analysis for Distributed Datasets With Orthogonal Integration Matrices

抄録

Advancements in Machine Learning (ML) are increasingly reliant on diverse datasets. However, combining multi-source data raises ethical concerns regarding user privacy and data misuse. This is further complicated by legal frameworks, like Japan's Act on the Protection of Personal Information, impacting ML deployment. Privacy-Preserving Machine Learning (PPML) addresses these challenges by ensuring data security, thereby supporting robust ML development. A key development in this area is the Data Collaboration (DC) framework, which facilitates secure ML training by integrating dimensionally reduced Intermediate Representations (IR) from distributed data. Current implementations face challenges with IR integration, affecting model stability. Our research presents an innovative enhancement to the DC framework, employing orthogonal integration matrices for IR integration. This solution aligns with the Orthogonal Procrustes Problem, offering an established analytical solution. Empirical assessments demonstrate that our approach notably improves recognition performance, surpassing traditional DC analysis methods. This study contributes to ML technologies' ethical and efficient advancement, respecting privacy concerns while optimizing data amalgamation.