Implementation guide
OWASP AISVS v1.01
Monitoring, Logging & Anomaly Detection · Training Data & Model Lifecycle Audit
Record dataset lineage
Verify that dataset lineage records each dataset and its components, including all transformations, augmentations, and merges.
OWASP AISVS v1.01, requirement 12.5.1
In plain English
Record the lineage of every dataset: what it is made of, and every transformation, augmentation and merge applied to it. When a source turns out to be poisoned, unlicensed or subject to a deletion request, lineage tells you which datasets and models are affected.
What good looks like
- Each dataset version has a lineage record naming its sources and their versions.
- Every processing step is captured, with the code version that ran it.
- You can trace a deployed model back to its source data, and a source forward to the models that used it.
- Lineage is recorded automatically by the pipeline.
How teams usually do it
- Data versioning and lineage tools in the pipeline, for example DVC, lakeFS, OpenLineage with Marquez, or your data platform's built-in lineage.
- Recording dataset hashes in the training run metadata and the model registry.
Suggested evidence
- Lineage for the training data of a deployed model.
- A worked trace from one source to every model that used it.
Common mistakes
- Lineage that stops at the first processed copy.
- Lineage reconstructed from memory during an audit.
- Leaving out synthetic or augmented data, which has sources of its own.
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.