I'll be interested in understanding the technical safeguards as well, since I am breaking my head on how to do this. One of my projects was about doing OCR on some confidential satellite images and we didn't want to outsource the OCR process solely because we couldn't figure out a scalable way to obfuscate data selectively in images. One early attempt was to recognize numbers using MNIST, get the boundary and clip that boundary with a similar sized black square. This doesn't work for all forms of text though.