LowModel Inversion○Theoretical
Model inversion / data reconstruction
Privacy & Data ProtectionDescription
Attackers reconstruct sensitive training inputs from model outputs or gradients.
Example scenario
Aggregated outputs let an attacker reconstruct features of confidential customer profiles.
Real-world evidence○Theoretical
While purpose limitation violations are a well-established GDPR compliance concern, documented enforcement actions that specifically identify an AI model as the mechanism by which data was processed beyond its collected purpose — as opposed to broader data governance failures — remain rare and are not yet a confirmed production AI incident category.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Differential privacy
- output perturbation
- access restriction
- gradient protection
- inversion testing.
Detection signals
Inversion attack simulation; reconstruction-fidelity testing.
Mitigating controls
4 Non-agentic controls