Privacy Configuration
The agent applies differential privacy to the quality metrics it reports, so only privacy-protected aggregates ever leave your site. The privacy parameters are configured on the Differential Privacy page of the agent UI and can also be set through environment variables. For a general introduction to how the framework protects data, see the Privacy & Security page.
Settings
| Setting | Description | Default | Validation |
|---|---|---|---|
epsilon | Privacy budget (ε). Smaller values mean stronger privacy but add more noise to results. | 3.0 | > 0 |
delta | Delta parameter (δ) — probability of privacy failure. Typically a very small value such as 1e-8. | 1e-8 | > 0 |
minThreshold | Minimum count threshold for low count suppression. Reported values below this are hidden. | 50 | >= 0 |
noiseMechanism | Noise distribution used. One of LAPLACE or GAUSSIAN. | LAPLACE | LAPLACE or GAUSSIAN |
Noise mechanisms
- Laplace: adds noise from a Laplace distribution. Best for unbounded queries and provides (ε, 0)-differential privacy.
- Gaussian: adds noise from a Gaussian (normal) distribution. Used for (ε, δ)-differential privacy where δ > 0. Provides better utility for bounded queries with the same privacy guarantee.
Gaussian constraint
When noiseMechanism is GAUSSIAN, epsilon must be <= 1.0. If you need a larger privacy budget, use LAPLACE.