The Discrete Gaussian Expectation Maximization (Gradient) Algorithm for Differential Privacy

Figure 1

Estimation error of GMM clipped vs. iteration t under different clipping threshold and budgets . (a) n = 1000; d = 20;  = 0.2, (b) n = 1000; d = 20;  = 0.5, and (c) n = 1000; d = 20;  = 1.

(a)

(b)

(c)