How to meet it
OWASP AISVS v1.01
Training Data Integrity & Traceability · Training Data Quality and Security Assurance
Defend against clean-label poisoning
Verify that defenses against clean-label poisoning attacks are implemented.
OWASP AISVS v1.01, requirement 1.3.5
In plain English
Defend against poisoning where the bad records look correctly labelled, so normal label checks will not catch them.
What good looks like
- The pipeline uses at least one defence aimed at clean-label poisoning.
- Data from less trusted sources is weighted or reviewed differently.
- The defence is tested with simulated attacks.
How teams usually do it
- Detection based on how training examples affect the model, such as spectral signatures or influence analysis.
- Training methods that limit how much any small group of examples can change the model.
Evidence an auditor would accept
- The chosen defence and why it fits your data.
- Results from a simulated clean-label attack.
Common mistakes
- Assuming label review covers this attack.
- Mixing data from unknown sources with trusted data without distinction.
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.