TA.Talal Alrawajfeh
← All research

Published · 2025

Noise-Aware Differentially Private Variational Inference

Bayesian inference that accounts for privacy noise introduced during variational optimization.

Authors
Talal Alrawajfeh, Joonas Jälkö, Antti Honkela
Venue
AISTATS 2025
Status
Published
Paper ↗Code ↗

Accounting for privacy noise in Bayesian inference

  1. Private optimisationVariational inference with differential privacy
  2. Noise-aware inferenceAccount for the privacy perturbation
  3. Posterior uncertaintyEvaluate coverage and predictive calibration
The method incorporates privacy noise into approximate Bayesian inference for high-dimensional, non-conjugate models.

This work studies how to account for privacy noise in differentially private variational inference, particularly for high dimensional, nonconjugate models. The research combines a mathematical formulation of noise aware inference with empirical evaluation on Bayesian regression tasks.

Published in the Proceedings of the 28th International Conference on Artificial Intelligence and Statistics (AISTATS 2025).