Implementation guide
OWASP AISVS v1.01
Training Data Integrity & Traceability · Data Labeling and Annotation Security
Redact sensitive data in labels
Verify that sensitive information in labels is redacted, anonymized, or encrypted before being used in any labeling artifact.
OWASP AISVS v1.01, requirement 1.2.3
In plain English
Remove or protect personal and confidential details before material goes to labellers, and keep them out of the label files. Labelling is often outsourced, so this is the point where sensitive data most easily leaves your control.
What good looks like
- Personal data is masked, replaced with realistic stand-ins or encrypted before items enter the labelling queue.
- Labellers see only the fields they need to make the judgement.
- Free-text notes and comments in label files are scanned for sensitive content.
- Images, audio and file metadata are covered as well as text.
How teams usually do it
- An automated detection and masking step in front of the labelling queue, for example Microsoft Presidio, Amazon Comprehend PII detection or Google Sensitive Data Protection.
- Labelling screens set up to hide fields the task does not need.
- For data too sensitive to mask well, an in-house labelling team working in a controlled environment.
Suggested evidence
- The redaction step, with test results showing how much it catches on a sample you checked by hand.
- A sample of labelling tasks as the labeller sees them.
- The contract terms that cover any outside labelling vendor.
Common mistakes
- Redacting the text and leaving names in attached images, filenames or EXIF metadata.
- Annotators pasting sensitive details into comment fields.
- Trusting automated detection completely. It misses things, so sample its output regularly.
Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.