Considerable research has been conducted on AI processing in cloud and edge systems. The purpose is to determine how to efficiently process big data in a virtual space and bring valuable information to the real space. One of the problems is how to perform AI processing while preserving the privacy of big data. To data, some learning methods have been proposed in this field. In particular, a machine learning method using secure distributed processing with divided data is one for realizing privacy and utilization. However, previous learning methods assumed that communication between servers is always possible. How should a learning method be realized when there is server that cannot communicate? In this paper, we show that effective BP methods, distributed systems with Q servers, can be achieved with accuracy comparable to traditional BP methods with 1/Q computation and communication complexity, when there are servers that cannot communicate.