Mechanisms
The mechanisms in this section extend the Adding Noise and Generalization categories of the masking-functions catalog with formal differential-privacy guarantees. Each page documents the SQL signature, when to use the mechanism, and known limitations.
If you're new to differential privacy, start with Concepts; the rest of this section assumes ε, sensitivity, and "local vs central" already mean something to you.
Catalog
| Mechanism | Output type | DP variant | Page |
|---|---|---|---|
| GRRM (Generalized Randomized Response) | categorical | local | GRRM |
| Laplace | numeric | local or central (means) | Laplace |
| Gaussian | numeric | local or central | Gaussian |
| One-hot Laplace / Gaussian | histogram bin vector | local | One-hot variants |
Decision matrix — which mechanism for which task?
Pick the row that matches what you're releasing, then the column that matches whether a curator can see raw values before noise is added.
| Releasing… | Local DP (untrusted aggregator) | Central DP (trusted curator) |
|---|---|---|
| One categorical value per row | GRRM | essentially random at any useful ε — release a histogram or aggregate instead |
| A frequency / histogram | GRRM + correction, or one-hot | direct count + Laplace on each bin |
| A numeric value per row | Laplace | rarely — central usually aggregates first |
| The mean of a numeric column | per-row Laplace then average | Laplace on the mean — far tighter |
Two parameters every mechanism takes
- ε (epsilon). The privacy budget. Smaller ε means stronger privacy and more noise. Composes additively across releases: a column released at ε=0.5 twice has spent ε=1.0 in total.
- Sensitivity. How much one row can change the unprotected query output. For per-row LDP it's the column's full public range; for a central mean it's the range divided by .
Where to next
- GRRM: the categorical mechanism.
- Laplace: numeric values, in both LDP and central DP forms.
- Gaussian: -DP with sub-Gaussian tails.
- One-hot variants: histogram release in a single LDP pass, no debiasing required.
- Post-processing: debiasing, confidence intervals, and mean estimation on GRRM output.