[3Rin4] Interactive 1

[3Rin4-38] Secure Deep learning that can build models on python

〇Ibuki Mishina¹, Koki Hamada¹, Shoko Nishida¹, Dai Ikarashi¹, Ryo Kikuchi¹ (1. NTT Secure Platform Laboratories)

Keywords:Secure Computation, AI, Deep Learning

There is a technology called SecureAI that performs machine learning while encrypting data.The use of this technology enables important data to be safely used for machine learning. However, SecureAI is still an application of cryptography, and there is nothing that can be used easily by users other than researchers of secure computing. Therefore, we are working on realizing a SecureAI library that is easy to use even for such users. In SecureAI, it is difficult to describe simply because it performs more complicated processing than the case without encryption. Therefore, in this paper, we implemented a Python library that automatically generates and executes a script for secure computation by describing and executing the processing in a format similar to the existing deep learning library. As a result, it became possible to safely build and learn models while encrypting the data using operations similar to those of existing libraries.