How to meet it
OWASP AISVS v1.01
Adversarial Robustness · Membership-Inference and Model-Inversion Mitigation
Calibrate the model's confidence
Verify that model outputs are calibrated to reduce overconfident predictions.
OWASP AISVS v1.01, requirement 11.2.3
In plain English
Calibrate output confidence so the model does not report itself as near-certain when it is not. Overconfident scores make some privacy attacks easier.
What good looks like
- Calibration is measured.
- Outputs are adjusted where calibration is poor.
- Calibration is rechecked after retraining.
How teams usually do it
- Temperature scaling or similar post-training calibration.
Evidence an auditor would accept
- Calibration measurements before and after.
Common mistakes
- Exposing raw probability scores that the model was never calibrated for.
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.