Application and evaluation of secure computation technology for the secure use of crop-breeding big data

Acceleration and efficiency improvement are necessary for the breeding of various crops that satisfy various needs, such as environmental adaptability, and for this purpose, the development of prediction and selection technology based on a large amount of breeding data (breeding big data) appears to be effective. To collect and utilize these breeding big data, a system is needed for the unified use of crop-breeding data held by various breeding organizations in Japan. However, there are many undisclosed breeding data for which intellectual property security, such as breeding registration, has not been carried out, and there are many difficulties in unifying and using these data on the assumption that they are publicly available. Thus, by applying secure computation technology that can encrypt and aggregate the data from multiple organizations and process and calculate them while maintaining confidentiality, the unified use of breeding data with high safety and convenience will become possible. For example, it will be possible to encrypt and collect sensitive breeding data held by multiple organizations, such as public laboratories, in each prefecture and seed companies without revealing the data to other organizations, create machine learning and prediction models with the encryption, and make predictions on the basis of the prediction models. Carrying out such integrated analysis using breeding data from multiple organizations will lead to acceleration and efficiency improvement of breeding. In this paper, the process from the registration of rice-breeding data in multiple cultivation areas to the prediction based on the machine learning model built using the breeding data for phenotype prediction was carried out on a secure computation system, and the applicability of the system was evaluated. Specifically, the prediction accuracy and processing performance were compared with the plain text, and confidentiality maintenance was evaluated in the whole process of learning. The results showed that the prediction accuracy was better when using a confidential calculation that can utilize the breeding data of multiple organizations while maintaining the secrecy of each organization’s data from the others in comparison to analysis that uses the data of a single organization. The results also showed that the data can be kept secret throughout AI processing of preprocessing, learning, evaluation, and prediction. Future work is to evaluate the usefulness in an actu