Implementation guide OWASP AISVS v1.01

Training Data Integrity & Traceability · Training Data Origin & Data Security

Protect training data in storage and transit

AISVS 1.1.3 Level 2

Verify that data integrity is provided when training data is stored and transferred.

OWASP AISVS v1.01, requirement 1.1.3

In plain English

Make sure training data cannot change unnoticed while it sits in storage or moves between systems. You should be able to show that the data a model was trained on is byte for byte the data that was approved.

What good looks like

  • When a dataset is approved, a cryptographic hash (SHA-256 or stronger) of each file is recorded in a manifest kept apart from the data.
  • Transfers between systems use TLS, and the receiving side checks the hashes on arrival.
  • The training job verifies the manifest before it reads any data, and stops if anything fails.
  • Approved datasets are stored so they cannot be overwritten in place.

How teams usually do it

  • A manifest of file hashes checked at the start of every run. Data versioning tools such as DVC or lakeFS do this for you.
  • Object storage with versioning and write-once retention (for example S3 Object Lock or Azure immutable blob storage) for approved datasets.
  • Signing the manifest, for example with Sigstore cosign, so nobody can swap the data and the hashes together.

Suggested evidence

  • Dataset manifests with hashes, and where they are stored.
  • Training logs showing the hash check ran and passed.
  • Storage settings showing versioning or retention locks on approved data.

Common mistakes

  • Recording hashes and never checking them before training.
  • Keeping the manifest in the same bucket as the data, with the same write permissions.
  • Approving a dataset in one location and training from a copy somewhere else, with nothing linking the two.

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.