2025 Theses Doctoral

Optimizing Privacy Budget Management in Differentially Private Systems

Kostopoulou, Kalliopi

Modern computing systems increasingly operate under stringent resource constraints—whether in the form of traditional hardware resources like CPU and memory, or novel, non-traditional resources such as user privacy.

This thesis explores systems and algorithmic techniques for efficient resource management in two distinct domains: (1) the emerging field of privacy-preserving data analytics, where privacy itself becomes a scarce and quantifiable resource to be allocated; and (2) distributed transaction processing, where lock-based contention and commit coordination determine throughput under high load.

In the first part of the thesis, we present a series of systems—DPack, Turbo, and Cookie Monster—that treat differential privacy budgets as consumable system resources. Each system targets a different layer of the privacy-preserving computing stack, from workload schedulers to database query caches to browser-based advertising measurement. Despite the diversity of applications, they all aim to improve the efficiency with which private data can be used, supporting more useful computation under fixed privacy guarantees.

The second part of the thesis shifts domains to distributed databases and introduces Sangria, an adaptive protocol that dynamically switches between conservative and pipelined commit strategies based on runtime conditions. Although this work is unrelated to privacy, it shares a common methodological theme: maximizing efficiency under contention and resource pressure. Together, these contributions illustrate the importance—and the diversity—of efficient resource allocation across modern computing systems, from privacy-aware data processing to classical transaction management.