The Japanese Society for Artificial Intelligence

In recent years, machine learning has made it possible to utilize various types of data. On the other hand, as the opportunities to use data increase, data breach from machine learning models has been pointed out. For example, consider a model that provides input candidates while writing an e-mail. If this model outputs a credit card number that exists in the training data, this is a data breach. In this research, as a countermeasure against data breach, we tackled the task of text generation by RNN using DP-Adam, an optimization algorithm that satisfies differential privacy. We tested whether it is possible to prevent the exposure of dummy data, which is assumed to be personal information. As a result, we confirmed that the DP-Adam model and the L1-regularized model avoided the dummy data exposure. However, the text generated by the L1-regularized model contained words that did not exist.