For those who are interested in machine learning and/or security, here is a research which discusses extracting secrets from machine learning systems:
The authors say deep learning models can memorize secrets which the training data may contain. Then the authors provide a couple of algorithms which were successfully used for extracting sensitive data from black box machine learning systems. The authors also show that unintended memorization is not the result of overfitting. Finally, they discuss several ways to mitigate the problem.
According to Mr. Schneier, there is a lot more research to be done here. So good luck :)