Organizing the Protection of Confidential Information Contained in Datasets when Training Neural Networks on Remote Electronic Computers and Cloud Services

Abstract

Various companies and researchers invest significant material and time resources to collect the necessary datasets, and as a consequence, wish to keep them secret. To analyze the collected datasets and further apply them in the training of neural networks for specific tasks requires the appropriate hardware, which not all machine learning researchers are able to acquire. To solve this problem, many IT corporations, such as Amazon or Google, provide access to their powerful hardware infrastructure (on a reimbursable and non-reimbursable basis) with high computing power for training neural networks. The semantics of the datasets on which the training of neural networks will take place is open. Therefore, there is a need to protect the semantics of the data on which neural networks will be trained in the cloud services or on remote computers