Speaker:

Apostolos Pyrgelis (RISE Research Institutes of Sweden)

Title:

Building Effective Privacy-Preserving Data Analytics with PrivaGym

Abstract:

Modern data analytics improve various aspects of our lives, e.g., healthcare, but they require access to large amounts of sensitive data raising significant privacy concerns. Hence, robust mechanisms are needed to safeguard data privacy while enabling meaningful analytics. However, deploying privacy defenses remains a fundamentally challenging problem as it requires carefully tuning hyperparameters that govern complex privacy-utility tradeoffs. This process lacks systematic technical guidelines, relies on manual and error-prone expert effort, and is computationally expensive. 

To address this critical gap, in this talk, we will introduce PrivaGym, an agentic environment designed to rigorously balance privacy and utility. In PrivaGym, a privacy agent interacts simultaneously with worst-case privacy adversaries and utility tasks to autonomously optimize the hyperparameter configurations of privacy defenses. We will model hyperparameter tuning as a pure exploration multi-armed bandit problem, enabling principled and data-driven identification of optimal configurations through rewards that explicitly capture the balance between privacy and utility. We will demonstrate via empirical experiments on various datasets and use-cases such as privacy-preserving statistics, clustering, and machine learning, that the privacy agent consistently and efficiently identifies optimal defense configurations, dynamically adapting to data characteristics, inference attacks, utility objectives, and end-user preferences, while outperforming baseline agents by one to two orders of magnitude.

Short Bio:

Apostolos Pyrgelis is a Senior Researcher at RISE Research Institutes of Sweden and a member of the Cybersecurity Unit. He received the PhD degree from University College London (UCL), United Kingdom, and the MSc and BSc degrees in Computer Engineering and Informatics from the University of Patras, Greece. His research interests include privacy-enhancing technologies, applied cryptography, and he enjoys working on problems at the intersection of machine learning analytics and security or privacy. More information can be found at his personal webpage (https://mex2meou.github.io/).