AIMGenerator<T>
Models & Types · AiDotNet.NeuralNetworks.SyntheticData
AIM (Adaptive and Iterative Mechanism) generator for differentially private synthetic data (McKenna, Mullins, Sheldon, Miklau 2022).
Inheritance: SyntheticTabularGeneratorBase<T> → AIMGenerator<T>
For Beginners
AIM works like a privacy-preserving census: it privately counts how many people fall into combinations of categories, choosing adaptively which combinations to count next, and then fits a probabilistic model that agrees with all those noisy counts.
How It Works
The algorithm follows the paper’s Algorithm 1 and the reference implementation in the private-pgm (mbi) library:
- The (epsilon, delta) budget is converted to a zCDP budget rho (Canonne, Kamath and Steinke 2020).
- Every one-way marginal is measured first with Gaussian noise of scale
sigma_0 = sqrt(T / (2 alpha rho)), T = 16 d rounds, alpha = 0.9.
- Each round selects a marginal from the downward closure of the workload with the exponential
mechanism (epsilon_t, starting at sqrt(8 (1 - alpha) rho / T)), using the quality
w_r (||M_r(D) - M_r(p_hat)||_1 - sqrt(2/pi) sigma_t n_r) and sensitivity max w_r. Only
candidates that keep the junction tree within MaxModelSize * rho_used / rho are eligible.
- The chosen marginal is measured with Gaussian noise of scale sigma_t, and p_hat is re-estimated by
Private-PGM: entropic mirror descent on the log-potentials of a graphical model whose cliques are the
measured marginals, minimizing sum_i ||mu_i - y_i||^2 / (2 sigma_i^2) with Armijo
backtracking.
- If the new estimate moved the chosen marginal by no more than
sqrt(2/pi) sigma_t n_rin
L1, the noise is halved and the selection budget doubled (sigma/2, 2 epsilon).
- When the remaining budget cannot fund two more rounds, the last round spends all of it.
Synthetic rows are drawn from the estimated graphical model by sampling attributes in reverse elimination order, each conditioned on its earlier neighbours in the triangulated model graph.
Constructors
| Constructor | Summary |
|---|---|
AIMGenerator(AIMOptions<T> options) |
Initializes a new AIM generator. |