Noise-Aware Differentially Private Variational Inference
The method introduces Noise-Aware DP Variational Inference (NA-DPVI) to handle differential privacy noise in approximate Bayesian inference, extending it beyond simple models to high-dimensional and non-conjugate ones. It uses stochastic gradient variational inference combined with a post-processing model to capture both DP uncertainty and data modeling uncertainty, achieving competitive performance on various models.