Question

A machine learning team outsources image-labeling to a third-party annotation workforce. The raw training images contain individuals' faces, license plates, and street signs. Privacy engineering is asked to build a data-oriented control that reduces privacy exposure to the annotators while preserving the labels the team actually needs (object bounding boxes for vehicles). Which technical approach best satisfies privacy by design in the annotation pipeline?

Accepted Answer

Automatically blur or redact faces and license plates in a preprocessing step before images are released to annotators, retaining only the visual content needed for the vehicle-labeling task

Answer

Require annotators to sign confidentiality agreements and complete privacy training before accessing the unmodified images

Answer

Encrypt the image files at rest in the annotation platform and rotate the encryption keys weekly

Answer

Log all annotator access to the images and review the audit trail monthly for anomalous viewing patterns