Federated and differentially private estimation of KL divergence

Abstract:Measuring distribution drifts is a key task in managing distributed, sensitive data, as it underpins a wide range of federated learning and analytics applications. In many practical settings, however, directly sharing such information is either undesirable (e.g., due to privacy concerns) or infeasible (e.g., due to high communication costs). In this work, we present FedPriKL, a novel method for estimating the KL divergence of data across federated computational models under differential privacy (DP) guarantees. We establish its theoretical properties, showing that FedPriKL is unbiased with low and bounded sensitivity and variance, thereby ensuring strong utility under DP. In addition, we present an empirical study that explores parameter choices to optimize accuracy while minimizing communication overhead. Our experiments demonstrate that FedPriKL achieves accuracy comparable to a similar sampling-based non-private estimator, while delivering greater stability and accuracy than a baseline variant in which users perturb their data locally to obtain privacy guarantees.

Machine Learning (cs.LG); Databases (cs.DB)

arXiv:2411.16478 [cs.LG]

(or arXiv:2411.16478v3 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2411.16478 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.