Ask HN: ML – if training data are private documents

Hi. I am working on building an application using ML on confidential finance documents. Data I require in huge number are private information which I can never have legally without the permission of the owner.

Has anyone used ML on such private documents, if yes what approaches did you take to collect such data?