Online Differential Privacy Bayesian Optimization with Sliced Wasserstein Compression

ICLR 2026 Conference Submission16063 Authors

Keywords: Online Learning, Bayesian Optimization, Local Differential Privacy, Sliced Wasserstein Distance, Kernel Compression

TL;DR: We propose a novel online differentially private Bayesian Optimization framework that enables zero-order optimization with rigorous privacy guarantees in dynamic environments.

Abstract: The increasing prevalence of streaming data and rising privacy concerns pose significant challenges for traditional Bayesian optimization (BO), which is often ill-suited for real-time, privacy-aware learning. In this paper, we propose a novel online locally differentially private BO framework that enables zero-order optimization with rigorous privacy guarantees in dynamic environments. Specifically, we develop a one-pass Gaussian process compression algorithm based on the sliced Wasserstein distance, which effectively addresses the challenges of kernel matrix scalability, memory efficiency, and numerical stability under streaming updates. We further establish a systematic non-asymptotic convergence analysis to characterize the privacy–utility trade-off of the proposed estimators. Extensive experiments on both simulated and real-world datasets demonstrate that our method consistently delivers accurate, stable, and privacy-preserving results without sacrificing efficiency.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 16063