Title:

DP Noise Calibration for Correlation-Resilient Privacy Guarantees

 

Abstract:

In this talk, we address a fundamental challenge in privacy-preserving data analysis: correlations between records can lead to privacy leaks that standard Differential Privacy (DP) fails to capture. We explore Bayesian Differential Privacy (BDP), an extension of DP that explicitly models such dependencies. Our goal is to examine whether BDP can be made practical in common data settings without incurring a significant loss in utility. We present our latest results on both general and structured correlation models, with a focus on Markov chains and multivariate Gaussian distributions. We establish theoretical connections between DP and BDP, introduce a method to adapt existing DP mechanisms to meet BDP requirements, and provide utility guarantees supported by empirical evaluation. Overall, our findings indicate that BDP can provide meaningful privacy protection for correlated data while maintaining competitive utility, moving it a step closer to practical use.

 

Bio:

Patricia Guerra-Balboa is a PhD student at the Chair of IT Security at KIT. She has a background in mathematics. Her current research interests lie in privacy-preserving data analysis with a strong focus on differential privacy and attack resilience.

 

Speaker:

Patricia Guerra-Balboa

PhD Student
Data Privacy Group
Chair of Privacy and Security
Karlsruhe Institute of Technology