howtosecure.ai
Implementation guide OWASP AISVS v1.01

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

Train only on the data you need

AISVS 1.1.1 Level 1

Verify that training data includes only features, attributes, and fields required for the model's stated purpose.

OWASP AISVS v1.01, requirement 1.1.1

In plain English

Train only on the fields the model needs for its stated purpose. Every extra column is one more thing the model can memorise and leak, and one more thing an attacker can poison. This is data minimisation, the same principle privacy law already asks of you, applied to training sets.

What good looks like

  • The model's purpose is written down in a sentence or two, and every field in the training set has a stated reason tied to it.
  • Direct identifiers (names, emails, account numbers) and special-category data are dropped or pseudonymised before training unless the purpose needs them.
  • The extraction step selects fields by name, so a new upstream column stays out until someone approves it.
  • The field list is reviewed whenever the purpose or the source system changes.

How teams usually do it

  • Start with a one-page data contract per training set: field, type, reason for inclusion, owner.
  • Build the dataset with an explicit column list in the query or pipeline code. Never use SELECT * against a production table.
  • For free text, run a personal-data detector (for example Microsoft Presidio or a cloud DLP service) and mask what it finds before the text enters the training store.

Suggested evidence

  • The purpose statement and the approved field list for each dataset.
  • The extraction query or pipeline code, showing that only those fields are selected.
  • A change record from the last time a field was added, with the reason and who approved it.

Common mistakes

  • Copying whole tables into the training store "in case they are useful later".
  • Removing names but keeping fields that identify a person in combination, such as postcode, date of birth and job title.
  • Forgetting free-text fields. A notes column often holds more personal data than every structured field put together.

Related requirements

1.1.2 Keep an inventory of training data sources · 1.2.3 Redact sensitive data in labels · 11.2.4 Train on sensitive data with differential privacy

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.