Toward the development of learning methods with distributed processing using securely divided data
Hirofumi Miyajima, Noritaka Shigei, Hiromi Miyajima, Norio Shiratori · Computers & Electrical Engineering · 2025
To pave the way to a super-smart society, artificial intelligence (AI) methods are being developed to discover and analyze necessary information instantly from cyberspace and utilize it in physical space. However, privacy protection is necessary for AI to process big data in cyberspace. From the viewpoint of developing safe and secure machine learning methods, research on (1) homomorphic cryptography, (2) differential privacy, (3) secure multiparty computation, and (4) federated learning is underway. The goal of these studies is to develop useful learning methods while maintaining data privacy. We propose a method to address the trade-off between security and usability in machine learning. This method balances usability and data confidentiality by using decomposed data to achieve secure distributed processing. However, such methods using distributed processing increase computational and communication overhead as the number of servers increases. To address this problem, we propose a method to control the computational complexity as the number of servers increases. On the basis of these studies, this study first systematically addresses the construction of secure distributed processing methods with decomposed data. A comprehensive approach is essential to advance the field and allow these methods to be effectively applied to different domains. On the basis of these methods, we propose back-propagation and neural gas learning methods with reduced computational and communication requirements. We then apply the proposed methods to numerical simulations of class classification and clustering problems and show that accuracy comparable to that of conventional models can be achieved with 1 / Q computational and communication complexity for distributed models with Q servers.