Researchers in the field of materials science are applying informatics methods to optimize the properties of materials. To accelerate their search, they need to gather data; however, they face the dilemma of whether to reveal or conceal their data. Secure computation offers an alternative; it allows the use of data without exposing them. We recently reported the development of an application that performs Bayesian optimization using secure computation. A statistical benchmark revealed its advantages even when only negative data were used.

We adopted a secure computation scheme using secret sharing, in which the data were decomposed into unintelligible shares. For example, consider the value of a physical property, say 137, and its decomposition into two unintelligible values: 137 = 878 + (-741). These unintelligible values, 878 and -741, are called shares. By distributing shares to two different servers, the value cannot be reconstructed unless both shares are obtained. Arithmetic operations can also be performed without the need for data reconstruction. In principle, a combination of these methods enables any type of calculation, albeit at the expense of additional computations and communication.

The aim of this study was to implement a practical method by devising an efficient approximation that can accelerate the search without compromising its efficiency. Although researchers have discussed efficient algorithms for Bayesian optimization using secure computation, there have been no reports on the implementation of a secure Bayesian optimizer.

In this study, we developed a practical and secure Bayesian optimizer. We demonstrated that the secure optimizer compares favorably with an ordinary optimizer in terms of the number of data acquisitions required. Sharing data has also been shown to significantly accelerate searches. Even if sharing is limited to negative data, it is beneficial for the search because it provides information about where the optimal system is unlikely to be found. We believe that this study demonstrates the potential of secure computation to open a new path in materials science and promote data sharing without exposing data.

(Written by Taro Fukazawa on behalf of all the authors)

Share this topic

Fields