How to meet it
OWASP AISVS v1.01
Training Data Integrity & Traceability · Training Data Origin & Data Security
Watermark your datasets
Verify that datasets are watermarked so their use can be attributed and any unauthorized use detected.
OWASP AISVS v1.01, requirement 1.1.5
In plain English
Mark datasets so you can show later that a given model or copy used them, and spot when they turn up somewhere they should not.
What good looks like
- High-value or licensed datasets carry a watermark or fingerprint that survives normal processing.
- There is a documented way to test a suspect model or dataset for the mark.
- Records show which watermark went into which dataset version.
How teams usually do it
- Inserting a small number of unique canary records that a model trained on the data would reproduce.
- Dataset fingerprinting tools that embed statistical marks.
Evidence an auditor would accept
- The watermarking method and the register of marks per dataset.
- Results from testing a model known to be trained on the data.
Common mistakes
- Watermarks that are removed by routine deduplication or cleaning.
- Marking the data with no plan for how you would ever check for the mark.
Not sure which requirements apply to your system? Take the assessment.
Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.