WeaknessASC-2026-0031Other weakness
Membership inference and model inversion
LowWorkarounds only
Attackers query a model to learn whether a record was in its training data, or to reconstruct sensitive attributes of training examples.
How to detect it
- High-volume queries probing confidence on specific records
- Models trained on sensitive data and exposed with detailed scores
Workarounds
- Return labels rather than raw confidence scores where possible
- Use differential privacy or aggregation when training on sensitive data
- Rate limit and monitor query patterns
Mainly a risk for models trained on sensitive records and exposed through an API.
How to investigate
- Review query logs for systematic probing
- Assess which training records could be exposed