SECURE MACHINE LEARNING SYSTEM BASED ON LIMITED ACCESS DATA FROM DIFFERENT SOURCES Russian patent published in 2024 - IPC G06F21/62 G06N20/00 

FIELD: physics.

SUBSTANCE: invention relates to a secure machine learning system based on limited access data from different sources. System comprises two embedded one into the other protection circuits, a blind data processing software-hardware complex (BDP SHC), located in an external protection circuit, outside of which there are external terminals, and a blind data processing computer (BDPC) located in the internal protection circuit, within the boundaries of which the protected data are processed in an open form, wherein all functional subsystems, apart from the machine learning subsystem which processes the protected data in the open form, are implemented in such a way that their parts, which process data of different degrees of criticality, are located on the structural elements of different levels of nesting of the circuits, and the machine learning subsystem is fully implemented within the internal protection circuit, wherein the system provides for the execution of computational processes of constructing a predictive machine learning model, its validation and application, including decrypting data using a supplier key, processing and encrypting data using a storage key, prepared for issuing as resultant reports, without disclosure of data combined from different sources, excluding acquaintance with them, copying and modification, but with possibility of their use for extraction of knowledge at the stage of machine learning, for which data are loaded into system from different sources in pre-encrypted form, after which in the system at the stage of machine learning a process of their automatic processing is performed, which includes a fixed sequence of data processing commands, which does not lead to acquaintance with them of personnel and to data leakage within the information system.

EFFECT: high level of data security.

4 cl